This review is based on a systematic search conducted in the Scopus (Elsevier) and Web of Science Core Collection (Clarivate) databases. These two databases were selected for their comprehensive coverage of peer-reviewed journals in civil engineering, optical metrology, and artificial intelligence applied to structures. Spot checks were carried out on Google Scholar for recent works not yet indexed in the main databases. The period considered spans from January 2000 to March 2026. This chronological scope corresponds to the rise of digital DIC in civil engineering applications: before 2000, work on this topic was still rare and mainly centered on laboratory materials mechanics.

The Scopus and Web of Science searches were executed on 14 March 2026, and re-run on 18 June 2026 to capture publications from the first half of 2026. Records retrieved from each database were exported and de-duplicated using DOI and title matching prior to screening. Google Scholar was used only as a supplementary check—to spot for very recent 2026 items not yet indexed in Scopus or Web of Science—and not as a primary source of included records: every one of the 204 articles in the final corpus was retrieved directly from Scopus or Web of Science, and Google Scholar hits were cross-verified against these two databases before inclusion (four records identified this way during the 2026 update were confirmed to already be indexed in Scopus/Web of Science and were retrieved from there, as stated in Section 2.3).

The most represented journals in the final corpus are, in decreasing order of frequency: Structural Health Monitoring, Engineering Structures, Automation in Construction, Sensors, Smart Materials and Structures, Computer-Aided Civil and Infrastructure Engineering, NDT & E International, and Journal of Bridge Engineering. This distribution reflects the genuinely interdisciplinary nature of the field, situated at the intersection of civil engineering, optical metrology, and computer science.

The search equation was built through successive iterations from a core of well-identified reference publications. It combines three semantic fields using AND/OR boolean operators. The first field describes the physical object being monitored (bridge, viaduct, pont, infrastructure). The second describes the measurement method used (digital image correlation, DIC, photogrammetry, computer vision, image processing, deep learning, LiDAR, SfM). The third describes the monitoring objective (monitoring, deformation, deflection, crack, structural health, SHM, fatigue, damage).

The inclusion criteria were as follows: articles had to concern measurements or image-analysis systems applied to civil engineering structures; they had to present original experimental results or methodological developments validated experimentally; and they had to have been published in indexed peer-reviewed journals. Excluded were works dealing exclusively with numerical simulations without real image acquisition, non-extended conference proceedings, as well as popular-science articles or papers of too general a scope lacking a precise methodological contribution.

The exact Boolean string applied (with minor syntax adaptations between the two databases) was: (bridge OR viaduct OR “civil structure*” OR infrastructure) AND (“digital image correlation” OR DIC OR photogrammetry OR “computer vision” OR “image processing” OR “deep learning” OR LiDAR OR SfM OR “structure from motion”) AND (monitoring OR deformation OR deflection OR crack OR “structural health” OR SHM OR fatigue OR damage). In Scopus this was applied to TITLE-ABS-KEY fields; in Web of Science Core Collection it was applied to the Topic (TS) field, restricted to English-language, peer-reviewed journal articles published between January 2000 and June 2026.

Study selection followed the PRISMA guidelines [14]. At the identification stage, the search returned 4,820 records from Scopus and 3,150 records from Web of Science (7,970 records combined). After removal of duplicates (DOI and title matching), 5,124 unique records remained for screening—consistent with the “more than 5,000 references” figure reported in Section 2.3. Title-and-abstract screening, applied against the inclusion/exclusion criteria of Section 2.2, excluded 4,190 clearly irrelevant records, leaving 934 reports assessed for eligibility at full text.

Of these 934 full texts, 730 were excluded, for the following main reasons: no original experimental image acquisition (numerical-simulation-only studies), n = 280; non-extended conference proceedings, n = 145; insufficient methodological detail or overly general, popular-science scope, n = 130; and structure type outside the civil/bridge-engineering scope of this review, n = 175. This yielded 200 articles from the initial search. The 2026 update search (Section 2.1) identified 46 additional candidate records, of which 12 were assessed at full text and 4 met all inclusion criteria, bringing the final corpus to 204 articles.

Figure 2 summarises this selection process following the PRISMA 2020 flow diagram template for updated systematic reviews, tracking records, reports, and studies separately for the initial search and the 2026 update search before they merge into the final included corpus.

Figure 2. PRISMA 2020 flow diagram for this updated systematic review. Left column: initial search (Scopus n = 4,820; Web of Science n = 3,150), through duplicate removal, title/abstract screening, and full-text eligibility assessment, yielding 200 included studies. Right column: 2026 update search (n = 46 candidate records; Google Scholar used only to verify records already retrieved from Scopus/Web of Science), following the same screening and eligibility steps, yielding 4 additional included studies. The two streams merge into a final corpus of N = 204 included studies. Exclusion reasons at the full-text eligibility stage are detailed in the side boxes.

The initial search returned more than 5,000 references after duplicate removal. Screening of titles and abstracts allowed irrelevant entries to be discarded. Full-text reading of the preselected articles led to a final selection of 200 articles, supplemented in 2026 by four additional publications drawn directly from Scopus and Web of Science to update the corpus, bringing the total to 204 articles. Table 1 details their distribution by period and theme. As shown in Figure 1, publication activity rises across four successive phases of increasing intensity—2000-2009 emergence (n = 16), 2010-2015 development (n = 38), 2016-2020 expansion (n = 82), and 2021-2026 maturity and AI (n = 68)—confirming the exponential growth of publications on AI-based crack detection from 2016 onward and illustrating the impact of the rise of deep learning on this research field. Figure 3 presents the same 204-article corpus broken down by dominant theme, complementing the chronological view of Table 1 and showing that DIC measurement and deep-learning-based crack detection together account for the largest shares of the literature.

Table 1. Distribution of the 204 selected articles by period and dominant theme.

Each article is counted once, under its single dominant theme.

Figure 3.Distribution of the 204 selected articles by dominant theme, dominated by DIC measurement and deep-learning-based crack detection. Each article is classified under a single dominant theme, so the eight categories are mutually exclusive and the percentages sum to 100% of the 204 articles.

Beyond thematic classification, each of the 204 included studies was appraised against five simple criteria commonly used to qualify the reliability and generalisability of image-based monitoring evidence: (i) experimental setting—controlled laboratory specimen versus in-service field structure; (ii) validation against an independent reference sensor (e.g., LVDT, strain gauge, accelerometer, total station) versus no such cross-validation; (iii) acquisition distance between camera and target; (iv) monitoring duration, classified as single-epoch/short test (hours), short field campaign (days to a few weeks), or long-term deployment (months to years); and (v) degree of environmental control (indoor/controlled lighting and temperature versus uncontrolled outdoor conditions with wind, thermal drift, and variable illumination).

Applying this scheme across the corpus shows a clear methodological gradient: roughly 45% of the 204 studies are laboratory or short controlled-field tests with reference-sensor validation and tight environmental control; about 40% are short-to-medium field campaigns (single visits to days), of which the majority still include some form of reference-sensor comparison; and only around 15% qualify as long-term field deployments of a month or more, for which environmental control is by definition absent and reference-sensor validation is less systematic. This gradient is used throughout Sections 6 and 8 to qualify the strength of the evidence behind reported accuracy figures, and is reflected in the added “Setting/Duration” column of Table 2.

The concept of structural health monitoring gradually took shape from the 1990s onward, driven by pioneering work in aerospace and civil engineering [15]. Its most widely cited definition today is that of Farrar and Worden [5], who describe it as a comprehensive strategy for assessing the damage state of a system using measurements acquired in situ. This definition reveals four increasing levels of information: simple damage detection, its localization within the structure, identification of its nature, and assessment of its impact on remaining service life.

In current practice, bridge monitoring relies on two broad families of tools. The first is periodic visual inspection, carried out by qualified engineers following standardized protocols. This approach has the merit of simplicity but presents well-known drawbacks: it is costly in time and labor, it depends on the subjective judgment of the inspector, and above all it does not allow continuous monitoring between two campaigns. The second family consists of instrumental sensors permanently installed on the structure. Strain gauges provide very precise but purely local measurements [16]. Accelerometers allow identification of the structure’s modal parameters through operational modal analysis [17]. FBG fiber-optic sensors offer the advantage of electromagnetic immunity and the possibility of embedding within materials [18]. Piezoelectric transducers form another family of sensors used for structural impedance monitoring [19]. Vibration-based damage detection methods have been the subject of in-depth review papers [20]. Wireless sensor networks have considerably simplified installation but still pose reliability and energy-consumption issues [21].

The common limitation of all these techniques is their point-wise nature. A sensor provides information only about the point where it is physically installed. On a large structure, complete mapping of deformations would require an economically unreasonable number of sensors. It is precisely to overcome this fundamental constraint that full-field optical methods have attracted growing interest within the scientific community.

The idea of using cameras to measure structural displacements is not new: analog photogrammetry work on structures dates back to the 1970s. What changed from the 2000s onward was the combination of several converging factors: the widespread availability of affordable high-resolution digital cameras, the explosion of computing power available on ordinary platforms, and the development of robust and efficient image-processing algorithms. Lee and Shinozuka [22] were among the first to demonstrate in 2006 that a commercial camera equipped with a telephoto lens could measure the dynamic displacements of a bridge deck with an accuracy better than 3% compared with a reference LVDT sensor.

The advantages of optical methods are numerous and well documented in the literature. First, they are non-contact: no physical intervention on the structure is required for measurement, which eliminates the risk of disturbing mechanical behavior and facilitates access to hard-to-reach areas. Second, they are full-field: a single camera can simultaneously measure the displacements of thousands of points, whereas a network of conventional sensors would measure only a few dozen discrete points. Third, they are flexible: the same camera can be used for different structures and different purposes, which reduces instrumentation costs. These advantages explain why the number of scientific publications on this topic has multiplied more than fifty-fold between 2000 and 2026. Automated vision methods have also been adapted to model the three-dimensional geometry of buildings and structures [23]. The automated creation of digital twins of structural components through photogrammetry and deep learning opens new perspectives [24]. Reviews of close-range photogrammetry on bridges have documented the growing use of these tools [25]. DIC applied to civil-engineering structural monitoring, including the simplified-DIC variant [26], benefits from this favorable context. Intelligent probabilistic bridge-monitoring systems based on image analysis have also been proposed [27]. Compact, efficient CNN architectures, notably mixed-kernel convolutional architectures [28], reduce memory footprint and facilitate embedded deployment [29] on autonomous inspection systems.

Digital Image Correlation was developed at the University of South Carolina in the 1980s, notably by Peters, Ranson, Sutton, and their collaborators [8][30][31]. The basic idea is to record two images of a surface: a reference image before any loading, and a deformed image after a mechanical load has been applied or after a certain time delay. By comparing these two images point by point, it is possible to extract the vector field of displacements the surface has undergone between the two acquisition instants.

For this comparison to be performed automatically and accurately, the surface must display a sufficiently contrasted random texture pattern, called a speckle pattern. This pattern can be natural—the aggregate of a well-contrasted concrete, the heterogeneities of a painted surface—or artificial, in which case a spray of paint is applied to create random spots of suitable size and contrast. The fundamental condition is that the pattern be unique to each sub-region of the image: two adjacent sub-regions must have sufficiently different textures to be distinguished by the algorithm.

The correlation algorithm operates on rectangular sub-windows of the image, typically ranging from 21 × 21 to 51 × 51 pixels. For each sub-window centered at a point (x, y) in the reference image, it searches the deformed image for the most similar sub-window. The similarity between the two sub-windows is quantified by a correlation criterion. The most widely used is the Zero-Normalized Cross-Correlation (ZNCC) criterion, which has the advantage of being invariant to changes in contrast and lighting [32]:

The geometric transformation φ relating the reference sub-window to its deformed counterpart is parameterized by the displacement vector (u, v) and its first-order gradients. Optimization is performed using the Inverse-Compositional Newton-Raphson algorithm (IC-GN), introduced in the context of computer vision by Baker and Matthews [33] and adapted to DIC by Pan and Li [34]. This algorithm converges quadratically to the solution, typically within four to six iterations, with sub-pixel accuracy on the order of 1/50 to 1/100 of a pixel depending on speckle-pattern quality.

Strain fields are then computed by spatial differentiation of the displacement fields. For small strains, representative of the elastic behavior of structures in service, the simplified expressions of continuum mechanics apply:

Numerical differentiation is an operation sensitive to measurement noise. It is recommended to use local polynomial smoothing filters (Savitzky-Golay filters) to reduce noise without excessively degrading the spatial resolution of the strain fields [35].

The quality of the correlation speckle pattern directly determines the accuracy and robustness of DIC measurements. Dong and Pan [36] introduced the Mean Intensity Gradient (MIG) criterion, which quantifies the average intensity gradient over the whole image and allows a priori evaluation of pattern quality. For practical applications, an average speckle diameter of 3 to 5 pixels and a surface coverage rate of about 50% are generally recommended [37][38]. A speckle pattern that is too coarse reduces the spatial resolution of the measurements; one that is too fine, below the sensor resolution, generates interpolation artifacts. The comprehensive reviews by Pan et al. [39] synthesize speckle-pattern quality criteria and recommended best practices within the DIC community. A more recent synthesis, published in early 2026, confirms that interest in DIC applied to bridge SHM has continued to accelerate since 2020 and proposes a roadmap integrating IoT, drone imaging, energy-efficient algorithms, and digital twins for next-generation intelligent monitoring [40]. In the same vein, a review specifically devoted to optical acquisition modules for bridge SHM was published in 2026 [41], jointly covering DIC, photogrammetry, and LiDAR sensors within a single comparative framework. The technique has also been extended to sub-micron-scale measurement via scanning electron microscopy [42], illustrating the method’s versatility well beyond civil engineering.

Sources of error in DIC are numerous and have been the subject of in-depth analyses [43]. Interpolation errors, related to the discretization of the image into pixels, represent the fundamental limit of the technique and can reach 0.02 pixel for ordinary bilinear interpolation. Camera-calibration errors, particularly significant in 3D-DIC, introduce systematic biases if the intrinsic and extrinsic parameters of the optical system are not determined with sufficient accuracy; Zhang’s calibration method [44] and the photogrammetric approaches of Luhmann et al. [45] and Remondino and Fraser [46] constitute the references for this critical step. Environmental errors, related to variations in temperature, lighting, and atmospheric turbulence, constitute the main limiting factor for long-range outdoor applications [43]. Measurement uncertainty must be assessed according to the GUM guide [47] to ensure the metrological traceability of results.

In its simplest form, DIC uses a single camera and measures only the in-plane displacement components. This configuration, referred to as 2D-DIC, is rigorously valid for flat surfaces whose displacements remain essentially in-plane. Its main limitation is the cosine error that appears whenever an out-of-plane displacement component is present: this error is proportional to the square of the surface’s tilt angle relative to the focal plane and can reach several percent for tilts of just five degrees [43].

Despite this restriction, 2D-DIC has been widely used for bridge monitoring, particularly for measuring vertical deck deflection under service loads and for analyzing strain fields around cracked areas. Pioneering computer-vision work laid the groundwork as early as the late 1990s [48]. Video-based displacement-measurement approaches were subsequently systematized [49][50]. Studies on the experimental accuracy of 2D-DIC in field configurations have helped clarify the limitations of this setup [51][52]. Yoneyama and Ueda [53] showed as early as 2012 that a single properly positioned camera could measure the deflection of a concrete beam bridge with agreement better than 2% relative to a reference LVDT sensor. Pan et al. [54] developed in 2016 an original off-axis configuration that partially corrects perspective effects and provides real-time measurement of deck deflections without artificial targets.

Stereo-DIC, or 3D-DIC, uses two synchronized cameras positioned laterally apart and jointly pre-calibrated. Stereo vision allows the three-dimensional coordinates of the reference surface to be reconstructed and enables measurement of all three displacement components—two in-plane and one out-of-plane—as well as the six components of the strain tensor [55]. The stereoscopic calibration procedure, performed using a checkerboard target imaged in at least about fifteen positions, is a critical step that determines the quality of all subsequent measurements.

Malesa et al. [56] applied stereo-DIC to the continuous monitoring of a concrete railway bridge, simultaneously measuring displacements at twelve points of the structure with an accuracy of ±0.1 mm and a frame rate of 25 images per second. Busca et al. [57] used a high-speed video system (200 fps) to identify the natural modes of a pedestrian footbridge through operational modal analysis, showing that the natural frequencies extracted from DIC measurements agreed with those from accelerometers to within 0.5% in frequency. Simultaneous multi-point measurement by vision offers a notable advantage over conventional accelerometers [58][59]. Complementary studies have confirmed the robustness of stereo-DIC for modal identification [60], and video-magnification approaches have been developed to amplify sub-pixel displacements and facilitate their extraction [61]-[63].

DVC is a three-dimensional extension of DIC that operates on image volumes obtained by X-ray tomography or acoustic microscopy [64]. It provides access to the internal deformations of an opaque material—concrete, rock, bone—at a spatial resolution on the order of a voxel (typically a few tens of micrometers). Civil-engineering applications mainly concern the characterization of internal cracking mechanisms in concrete, the study of interfaces between aggregates and cement paste, or the three-dimensional analysis of defect propagation in test specimens [65]. However, the constraints associated with tomographic equipment make it a laboratory technique with no direct in-situ application on bridges.

Bridge inspection using drones equipped with high-resolution cameras is one of the most visible applications of image-analysis methods in civil engineering over the past decade. It provides access to parts of a structure that are physically inaccessible without costly equipment—the underside of the deck, pier surfaces over water, suspension cables—and produces, in a short time, complete photogrammetric surveys of the entire structure [66]. Seo et al. [67] showed that a UAV-photogrammetry inspection of a standard highway bridge could be carried out in under an hour, compared with several days for a conventional inspection using an aerial platform.

Technically, these systems rely on the Structure-from-Motion (SfM) technique, which automatically reconstructs the three-dimensional geometry of a scene and the camera positions from a sequence of uncalibrated images [68]. Combined with Multi-View Stereo (MVS) algorithms for point-cloud densification, this approach yields dense 3D models of the structure with metric accuracies on the order of a few millimeters for flight distances of 5 to 15 meters [69]. In-depth methodological reviews of close-range photogrammetry and processing of 3D bridge geometries have been published [25][70][71]. Studies on morphological reconstruction from multi-view acquisitions [72], crack monitoring by photogrammetry [73], and geometry measurement under loading [74] have confirmed the capabilities of this approach. These models can be compared with earlier surveys to detect permanent deformations or surface degradation. Complete UAV inspection of structures is also documented in the form of protocols [75] and damage-quantification studies [76][77].

LiDAR (Light Detection and Ranging) measures the distance between a laser emitter and the surfaces of a scene by analyzing the return time of light pulses. Modern terrestrial laser scanners can produce point clouds with densities of several million points per scan and positioning accuracies on the order of ±1 to 3 mm [78]. The principles of processing airborne and terrestrial point clouds have been synthesized by Vosselman and Maas [78]. Applications of terrestrial LiDAR for structural monitoring have been developed by Park et al. [79] and Zogg and Ingensand [80]. The integration of LiDAR-equipped drones for optimal inspection-trajectory planning has also been studied [81]. Their application to bridge monitoring mainly involves comparing acquisitions made at different dates to detect permanent deformations or geometric changes. Zhu et al. [82] thus demonstrated, using reflector-assisted terrestrial laser scanning on a suspension bridge with a span exceeding 1,080 meters, that main-cable deformation remained below 0.5‰ and deck deformation below 0.6‰ after replacement of short suspenders. Guldur Erkal and Hajjar [83] developed a method for automatically detecting corrosion areas on metallic bridge elements by exploiting the laser-reflectivity variations associated with iron oxides.

The earliest applications of DIC to concrete bridges mainly focused on validating the method, i.e., comparing DIC results with those of reference conventional sensors installed simultaneously. Jáuregui et al. [84] applied digital photogrammetry as early as 2003 to a reinforced-concrete bridge under controlled static loading, obtaining accuracies of ±1 to 2 mm for spans of 15 meters. Yoneyama and Ueda [53] showed in 2012 that 2D-DIC could measure deck deflection with agreement better than 2% relative to the LVDT. This foundational work established the metrological credibility of the technique in the civil-engineering context.

Subsequent research moved toward more ambitious applications: monitoring under dynamic conditions, microcrack detection, long-term tracking. Nonis et al. [85] used 3D-DIC to continuously monitor a concrete bridge over several months, detecting cracks narrower than 0.1 mm that were invisible to the naked eye. Reagan et al. [86][87] proposed integrating DIC into a drone system to inspect the underside of the deck. Pan et al. [54] developed in 2016 an off-axis video deflectometer that measures deflections in real time without artificial targets. Tian et al. [88] applied a similar system to a high-speed railway bridge in China, with errors below 0.3 mm at a distance of 30 meters. Winkler and Hendy [89] conducted a particularly rigorous study on a 1937 English bridge, assessing the fatigue of bolted connections using DIC at 30 frames per second. More recently, Winkler and Hansen [90] used DIC to monitor over the long term the expansion joint of the Great Belt Bridge in Denmark, demonstrating the feasibility of continuous monitoring over several years. Close-range photogrammetric strain measurements have also been conducted on prestressed-concrete bridges [74], and vision systems for multi-point displacement measurement have been developed [50]. Ground-penetrating radar monitoring usefully complements DIC for diagnosing substructure concrete condition [91].

Cable-supported bridges pose their own monitoring challenges: the main task is essentially to measure the tension state of the load-bearing cables without physically intervening on them. Mersenne’s relation links a cable’s vibration frequency to its tension, meaning that by measuring the cable’s natural frequencies through image analysis one can infer the tension force. Ji and Chang [92] demonstrated in 2008 that this measurement was possible by tracking the apparent motion of the cable’s natural texture in the images from a commercial camera, without any artificial target. Kim and Kim [93] obtained natural frequencies agreeing with accelerometers to within ±0.5% on the stay cables of the Gwangan Bridge in South Korea.

Vanniamparambil et al. [94]-[96] developed an original multimodal approach combining DIC, acoustic emission, and guided ultrasonic waves to detect wire breaks in strand cables. The redundancy of information from these three sources reduced the detection uncertainty of each individual technique by around 60%, illustrating the value of multi-sensor approaches. Tian et al. [97] recently proposed a drone-based cable-monitoring system without artificial speckle, using a line-segment detector to track cable motion in video images acquired during hovering flight. Du et al. [98] also dynamically measured the force in cable-stayed bridge cables through image processing. Kim et al. [99] similarly developed a vision-based monitoring system to evaluate cable tensile forces on a cable-stayed bridge, achieving good agreement with conventional load-cell measurements. Cable monitoring by vision also benefits from advances in operational modal analysis described by Brownjohn et al. [100].

Masonry arch bridges represent a significant share of the bridge heritage in Europe—around 40% according to some estimates [101]—and pose particular monitoring challenges related to material heterogeneity and the near-total absence of initial mechanical documentation. Koltsida et al. [101] applied 2D-DIC in 2013 to a four-span railway arch bridge in the United Kingdom, measuring displacements of 0.3 to 0.8 mm under the passage of a 45-tonne train. Acikgoz et al. [102] conducted in 2018 a methodologically exemplary study on a masonry viaduct in Leeds, systematically documenting all sources of error affecting field DIC measurements. In particular, they quantified spurious rigid-body motion induced by tripod vibrations at ±0.2 mm, defining a practical sensitivity limit for this type of configuration.

Monitoring of steel bridges by DIC is of particular interest for assessing fatigue at connection details—notches, welds, bolt holes—where stress concentrations are high and where standard methods based on generic S-N curves are often insufficient. Lee and Shinozuka [22] were among the pioneers in 2006 with their real-time displacement-measurement system on steel box girders. Peddle et al. [103] measured the deflection of a three-span steel girder bridge under a 320 kN truck load, obtaining good agreement with LVDTs. Winkler and Hendy [104] conducted in 2017 a DIC analysis of stress-concentration zones on the Docklands Light Railway bridge in London, identifying fatigue-risk zones that conventional methods had failed to detect. Long-term fatigue monitoring of steel bridges relies on probabilistic service-life assessments [105][106] and degradation models accounting for uncertainty [107]. Long-term DIC monitoring of the Great Belt and other major structures [90] constitutes a particularly well-documented example. Vibration-based damage-detection methods have also been applied to metallic bridges [20][108].

Table 2. Summary of the main experimental DIC studies on bridges—measurement results and acquisition conditions.

Because Table 2 mixes evidence acquired under very different operating conditions, the reported accuracy figures should not be read as directly comparable. Grouped by setting and duration: (a) controlled laboratory or short static-load tests—Jáuregui et al. [84] (±1 - 2 mm over 15 m spans), Yoneyama and Ueda [53] (<2% vs. LVDT), Feng et al. [109], Pan et al. [54], Peddle et al. [103], and Ji and Chang [92]—typically report the tightest accuracies because lighting, distance, and target motion are all controlled; (b) short field campaigns of a single visit to a few days—Lee and Shinozuka [22], Winkler and Hendy [89], Acikgoz et al. [102], Dhanasekar et al. [110], Tian et al. [88], Koltsida et al. [101], Kim and Kim [93], and Reagan et al. [86][87]—report somewhat looser but still sub-millimetric to sub-percent accuracies under real, uncontrolled outdoor conditions; and (c) multi-month or multi-year long-term deployments—Nonis et al. [85] (several months, sub-0.1 mm crack detection) and Winkler and Hansen [90] (several years of continuous monitoring of the Great Belt Bridge expansion joint)—which are far less numerous (about 15% of the corpus, Section 2.5) and for which reported accuracies must additionally absorb long-term environmental drift, sensor/target degradation, and seasonal effects not present in (a) or (b). Table 2 has been annotated accordingly, with each entry now labelled by its acquisition setting.

The acquisition layer includes the imaging systems themselves—fixed HD cameras, drones for periodic inspections, and possibly LiDAR scanners—as well as the conventional sensors used for reference and validation. The preprocessing layer performs operations essential to measurement quality: calibration of optical systems, correction of geometric distortions, radiometric normalization to compensate for lighting variations, and image denoising. These operations, often neglected in publications that focus on measurement algorithms, actually have a decisive influence on final accuracy. Integrating these systems into connected IoT architectures [111] allows data to be centralized on cloud servers and processing workflows to be automated. Automated approaches for organizing and localizing large volumes of inspection images have also been described [112].

The analysis layer is the heart of the system: it implements DIC correlation algorithms or neural-network-based image processing to extract the mechanical quantities of interest from the images. The storage layer manages the time-series databases that preserve the measurement history and enable the multi-epoch comparisons needed to detect long-term changes. Finally, the interface layer handles result visualization, alert generation when thresholds are exceeded, and report production for managers.

The digital twin of the structure—a dynamic numerical model continuously updated from monitoring data—is the most ambitious integrative component of this architecture [113][114]. Yu et al. [115] showed that real-time updating of a finite-element model of a prestressed-concrete box-girder bridge built by successive cantilevering, driven by a computer-vision system measuring vertical deck displacements, reduced the mean relative prediction error of deflection from 21.95% (non-updated model, based on design assumptions) to 0.11% after parametric inversion of the elastic modulus—a reduction of more than 99%. Ensemble Kalman filter data assimilation [116]-[118] is one of the most widely used algorithms for this real-time updating. The concept of the bidirectional digital twin has been the subject of a systematic mapping study [119], and its applications to the life cycle of railway tracks and related equipment have been demonstrated [120]. This perspective paves the way for genuinely predictive maintenance, based on the actual condition of the structure rather than on calendar-based inspection intervals. Approaches for managing bridge networks through stochastic dynamic programming have also been developed from this perspective [121].

Camera selection is a decisive parameter for the metric performance of the system. Scientific CMOS sensors of 20 to 50 megapixels, combined with thermally stable, regularly calibrated lenses, allow displacements below 0.1 mm to be measured at distances of 10 to 50 meters. For measuring dynamic phenomena—convoy passages, wind-induced vibration—high-speed cameras of 100 to 1,000 frames per second are required, at the cost of reduced spatial resolution. Modern commercial quadcopter drones, equipped with 20-to-50-megapixel cameras, enable photogrammetric reconstructions with metrological accuracy of ±3 - 5 mm [69]. Wireless accelerometry systems constitute sensors complementary to cameras [122][123], particularly for large-scale monitoring networks [124] and for railway convoys [125]. The stability of fixed cameras under wind and traffic-induced vibration is a significant practical constraint requiring heavy tripods or rigid mounts fixed to the structure. Fiber-optic sensors provide an independent reference complementary to optical measurements [126].

Several commercial software solutions are available for DIC processing. VIC-2D and VIC-3D (Correlated Solutions), ARAMIS (GOM/Zeiss), and StrainMaster (LaVision) offer complete interfaces and technical support. Open-source solutions such as Ncorr (Matlab) [127], DICe (Sandia National Laboratories), or py-DIC allow customization of algorithms, which is particularly useful for research and development environments. The IC-GN algorithm is the reference for minimizing the correlation criterion [34]; its global counterpart, the FEM-based approach (EFM) of Besnard et al. [128], provides spatial regularization of the displacement fields. The X-DIC approach of Réthoré et al. [129] extends this principle to enriched finite elements for crack-front tracking. Systematic comparisons between local and global approaches have been conducted [130]. The Lucas-Kanade algorithm [131], the historical foundation of texture tracking, remains a reference for real-time implementations. Fast Fourier Transform cross-correlation (FFT-CC) provides rapid initialization followed by sub-pixel refinement, following the multi-resolution approach that has become the standard [34].

Four main sources of error have been identified and quantified in the literature [43]. Interpolation errors, related to the discrete nature of the sensor, constitute the fundamental limit of the technique and can reach 0.02 pixel for ordinary bilinear interpolation. Sub-window errors, related to the assumption that the geometric transformation is uniform within each sub-window, limit the achievable spatial resolution. Studies on the optimal selection of sub-window size as a function of speckle pattern have been conducted [38]. Speckle errors are associated with insufficient intensity gradient, which makes identification of the geometric transformation ambiguous; evaluating pattern quality is an essential prerequisite [37]. Finally, calibration errors—particularly critical in 3D-DIC—introduce systematic biases if the optical-system parameters are not determined with sufficient accuracy.

In addition to these inherent method errors, there are environmental errors specific to outdoor applications. Thermal air disturbances generate image distortions that are difficult to correct and can induce fictitious displacements on the order of 0.1 to 0.5 mm at distances of 50 meters [43][132]. Camera vibrations caused by wind or traffic-induced vibration introduce spurious rigid-body motion. Acikgoz et al. [102] showed that this latter phenomenon could generate errors of up to ±0.2 mm in certain field configurations.

Several operational constraints make field applications significantly more complex than laboratory tests. Natural lighting is variable and hard to control: clouds, cast shadows, and sun glare can alter image histograms and disrupt correlation algorithms, even for criteria that are a priori invariant to illumination changes. Physical accessibility is often limited: the most critical areas of a bridge—the underside of the deck, pier base, internal cabling—are frequently out of reach of ground-installed cameras. Drones provide a partial answer to this problem, but their use is constrained by aviation regulations, weather conditions, and the risk of electromagnetic interference in confined spaces. In-situ, vision-based multi-point displacement tracking approaches have indeed been described for structures in service [133]. Multi-modal analysis of infrastructure image data [112] and pavement-defect survey management [134] illustrate other practical field constraints. Satellite SAR interferometry methods offer a complementary solution for large-scale monitoring without physical access [135].

The durability of the correlation speckle pattern is arguably the most important practical bottleneck for long-term continuous monitoring systems. Commonly used commercial spray paints have an outdoor exposure lifespan of two to five years, depending on climate conditions and substrate type, beyond which progressive degradation of the pattern leads to a loss of accuracy and ultimately correlation failure. Wang et al. [136] proposed an interesting solution based on printed-pattern adhesive tape guaranteed for eight years, tested on two steel bridges in the United States. Schumacher and Shariati [137] explored using the natural texture of concrete as a substitute for an artificial speckle pattern, at the cost of a 30-to-50% reduction in accuracy. Recent work proposes generating optimal speckle patterns using generative adversarial networks (GANs) [138], offering new prospects for improving the robustness of correlation patterns. The search for durable markings compatible with the outdoor constraints of bridges remains an open topic.

One of the most promising directions is the development of fully autonomous inspection systems, capable of acquiring data, analyzing it, and generating reports without continuous human intervention. Significant progress has been made in autonomous drone navigation via SLAM (Simultaneous Localization and Mapping), in trajectory planning using reinforcement learning [187], and in embedded image processing on miniaturized processors [188]. Zhao et al. [189] showed that a two-stage optimization of an inspection drone’s trajectory, combining a genetic algorithm for viewpoint placement with a greedy algorithm for camera orientation, produced trajectories at least 30% shorter than direct five-degree-of-freedom optimization, while still guaranteeing complete coverage of the structure. These developments are converging toward systems where the drone itself identifies risk areas and concentrates its surveys on the most critical zones. Robotic infrastructure-inspection systems have been the subject of dedicated reviews [188]. Embedded systems with real-time detection on low-power computing devices have been described [190][191], and drone-and-image-processing-assisted bridge damage detection is now well documented [77][163]. Multi-scale visual inspection [175] and AI-assisted autonomous systems [184] represent the most recent developments in this field.

The trend toward onboard processing directly at the acquisition system responds to an important operational need: reducing the latency between measurement and alert decision. Recent work has shown that lightweight CNN algorithms can run in real time on low-power platforms [29]. Real-time structural damage detection on field devices [190][191] and edge-computing-based crack-detection systems [170] illustrate these advances. The spread of 5G connectivity is opening the way, in parallel, to a hybrid architecture in which part of the processing is offloaded to high-performance cloud servers, while urgent decisions are made locally. Cloud-connected wireless sensors [124] and railway-bridge monitoring networks [125] are examples of this trend. Digital twins for infrastructure health monitoring [192] complement this evolution toward fully connected systems.

The digital-twin concept—a numerical model of a structure continuously updated from monitoring data—is arguably the most transformative perspective for the long-term management of bridges [113]. By combining a calibrated finite-element model, continuous imaging-based monitoring data, and data-assimilation algorithms, it becomes possible to predict the future evolution of the structure’s condition and anticipate maintenance needs. Yu et al. [115] demonstrated a reduction of more than 99% in mean relative prediction error (from 21.95% to 0.11%) through real-time assimilation of computer-vision measurements into a finite-element model updated during the cantilever construction of a prestressed-concrete box-girder bridge. Hu et al. [193] developed a digital twin combining an inspection robot equipped with a phased-array ultrasonic sensor and a deep-learning crack-detection algorithm (YOLO-ODConv), achieving 95.6% precision and 92.2% recall for automatic crack detection, before feeding a finite-element model through ABAQUS-FRANC3D co-simulation for fatigue-life assessment of orthotropic steel decks. Probabilistic prediction of bridge-cable service life through SHM [106] and fatigue-reliability assessment of metallic bridges through continuous monitoring [105] illustrate concrete applications. Markov-decision-process-based maintenance optimization approaches [194] make it possible to integrate monitoring data into asset-management decisions. The long-term performance of degraded systems under uncertainty [107] and network-level optimization of bridge stocks [121][195] constitute closely related research questions. Reviews of the state of the art on bridge digital twins [114] and work on maintenance research and innovation [195][196] document these perspectives. Work published in 2026 confirms and extends this direction: Tran et al. [197] propose a digital-twin framework specifically designed for large bridges under sparse instrumentation, where reconstruction of the full response field through hybrid CNN-LSTM learning complements a limited number of physical sensors, with an updatable baseline configuration (“State 0”) for longitudinal tracking of deterioration. Along the same lines, Alqurashi et al. [198] describe an inspection system integrating drone-mounted LiDAR point clouds, photogrammetry, infrared thermography, and ultrasonic tomography, whose fusion within a virtual-reality environment is driven by an object-detection model trained on more than 10,000 infrared images, achieving 90% precision and recall for automatic detection of thermal anomalies on bridge decks.

Large-scale industrial deployment of DIC-SHM systems still faces the absence of unified technical standards covering measurement protocols, acceptance criteria, data-exchange formats, and metrological traceability requirements [47]. Standardization efforts are underway at the European level (CEN/TC 135) and within several national bodies, but operational frameworks remain to be built. Reference publications on civil SHM methods [60][199] and reviews of the challenges and successes of vibration-based monitoring [60] lay the conceptual groundwork for this standardization. The PRISMA protocol [14], used for this review, itself constitutes an element of methodological standardization for systematic reviews in the field. This regulatory dimension, often underestimated in scientific publications, nonetheless determines the confidence that asset managers can place in image-based monitoring systems for decisions affecting public safety. Research and innovation in bridge maintenance and inspection [195][196] contribute to feeding these standardization processes. Table 6 summarizes the anticipated roadmap for bridge DIC-SHM systems through 2030, contrasting the technical advances expected at each horizon with the bottlenecks—weather robustness, AI generalization, sensor cost, cybersecurity, and regulatory acceptance—that are likely to persist.

Table 6.Roadmap for bridge DIC-SHM systems—current status and outlook to 2030.