Drive-by Bridge Inspection

• Drive-by bridge inspection is a remote monitoring method that leverages vehicle vibrations, UAV imagery, and crowdsourced data to assess bridge conditions without direct sensor installations.
• It employs indirect structural health monitoring using techniques like frequency domain decomposition, unsupervised learning, and domain adaptation to reliably identify damage across diverse bridges.
• The approach integrates vision-based and digital-twin frameworks to enhance real-time maintenance decisions, scalability, and cost-effectiveness in bridge condition assessments.

Drive-by bridge inspection is a family of bridge condition assessment methods that infer structural condition without direct installation of dedicated sensors on the bridge, or that acquire inspection data from mobile or remote platforms rather than fixed bridge-mounted instrumentation. In the most specific usage, it denotes indirect structural health monitoring based on the coupled response of a passing vehicle and bridge; in broader recent usage, it also includes camera-based, UAV-assisted, and robotic inspection pipelines that exploit existing traffic cameras, public weather feeds, or autonomous platforms for rapid, low-cost, and scalable monitoring (Hurtado et al., 1 Oct 2025, Liu et al., 2020, Balijepalli et al., 14 Mar 2026).

1. Core concept and problem setting

Drive-by bridge inspection is motivated by the cost, logistics, and limited coverage associated with conventional direct instrumentation and periodic manual inspection. Vehicle-based approaches emphasize that bridge health can be monitored using the vibrations of drive-by vehicles, with benefits including low cost and no need for direct installation or on-site maintenance of equipment on the bridge (Liu et al., 2020). A related line of work frames this as Indirect Structural Health Monitoring (ISHM), using vehicle-mounted sensors to assess bridge condition without requiring direct instrumentation, and reports field validation on two full-scale bridges with a purpose-built electric inspection vehicle (Hurtado et al., 1 Oct 2025).

The sensing substrate varies by modality. In indirect SHM, the primary observable is the coupled vehicle-bridge response, often measured by accelerometers mounted on vehicle suspensions or chassis. In crowdsourced variants, smartphone-vehicle-trip data supply acceleration, GPS position, and orientation during ordinary crossings, enabling extraction of bridge dynamic characteristics from widespread mobile devices (Cronin et al., 2022). In hybrid digital-twin formulations, existing traffic cameras are repurposed as sensor feeds, and weather APIs provide environmental deterioration drivers, reducing reliance on dedicated sensor installations (Balijepalli et al., 14 Mar 2026).

A broader operational interpretation also appears in the visual-inspection literature. UAVs are used to collect high-resolution imagery from difficult-to-reach bridge regions, and autonomous ground or climbing robots carry cameras and NDE sensors across bridge surfaces. These systems do not all satisfy the narrow indirect-SHM definition, but they share the drive-by objective of remote, repeatable, and reduced-risk inspection from mobile platforms rather than fixed bridge instrumentation (Phan et al., 2024, La et al., 2017, La, 2017).

2. Indirect structural sensing from passing vehicles

In vehicle-based drive-by inspection, the central signal is the coupled vehicle-bridge interaction. One recent framework uses a custom-designed electric inspection vehicle with four suspension-mounted accelerometers and four load cells, autonomous constant-speed control, and repeatable path following by magnetic guidance. In that study, Frequency Domain Decomposition (FDD) was applied to the cross-power spectral density matrix

$$\mathbf{S}(f) = \mathbf{U}(f) \mathbf{E}(f) \mathbf{U}^H(f),$$

and drive-by measurements recovered bridge natural frequencies consistent with direct instrumentation on two full-scale bridges (Hurtado et al., 1 Oct 2025).

Unsupervised learning is increasingly used to convert spectral features into damage-sensitive indicators. A recent optimization study models the inspection vehicle as a half-car system coupled to an Euler-Bernoulli bridge, computes front and rear axle spectra, forms a residual spectrum $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$, and trains an adversarial autoencoder on healthy data. The damage index is the reconstruction error

$DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$

while separability between healthy and damaged states is quantified by the first-order Wasserstein distance

$$W_1(P_1, P_2) = \inf_{\gamma \in \Gamma(P_1, P_2)} \int |x_1 - x_2|\, d\gamma(x_1, x_2).$$

That work reports that vehicles with frequency ratio $0.3 \le \beta \le 0.7$ relative to the bridge first natural frequency are most effective, that the best value found was $\beta \approx 0.42$, and that vehicles near resonance perform poorly (Hurtado et al., 3 Oct 2025).

Crowdsourced smartphone sensing extends the same principle to opportunistic traffic. A statistical signal-processing pipeline maps each acceleration record $a_i(t)$ through a Synchrosqueezed Wavelet Transform to $T_{a_i}(f,t)$, then to spatial coordinates $T_{a_i}(f,x)$, and estimates an absolute mode shape from spatially aggregated amplitudes: $$B_{ij}=T_{a_i}([f_k-\epsilon/2,f_k+\epsilon/2], [j\times S, \max\{L_{br},\Delta + j\times S\}]),$$

$$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$0

The method was demonstrated on four bridges with span lengths ranging from about 30 to 1300 meters, with reported MAC values of 0.94 or greater relative to fixed-sensor references (Cronin et al., 2022).

These results suggest two distinct but complementary trajectories within indirect SHM. One trajectory seeks higher repeatability through purpose-built inspection vehicles and controlled operating conditions; the other seeks scale through crowdsourced smartphone data. A plausible implication is that future systems may combine both, using dedicated vehicles for calibration-rich campaigns and passively collected fleet data for temporal densification.

3. Knowledge transfer, domain adaptation, and multi-task diagnosis

A persistent obstacle in drive-by diagnosis is bridge-to-bridge distribution mismatch. Supervised models trained on one bridge often degrade when directly applied to another because vibration-response distributions differ across bridges. To address this, MT-DANN introduces a Multi-Task Domain Adversarial Neural Network for damage detection, localization, and quantification using vehicle vibration responses (Liu et al., 2020).

MT-DANN couples a shared feature extractor $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$1, task predictors $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$2 and $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$3, and a domain classifier $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$4 trained through a gradient reversal layer. Its objective combines task losses and domain loss: $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$5 The reported average accuracies on target bridges with no labels used in training were 94% for damage detection, 97% for localization, and 84% for quantification within one damage severity level (Liu et al., 2020).

HierMUD generalizes this strategy to hierarchical multi-task unsupervised domain adaptation. It distinguishes task-shared and task-specific features, retaining shared features for easier tasks such as detection and localization while adding task-specific feature extractors and domain classifiers for harder tasks such as quantification. The framework is grounded in a multi-task UDA generalization-risk analysis and uses a soft-max prioritized divergence term in the loss. Reported performance averages were 95% for damage detection, 93% for localization, and up to 72% for quantification, with roughly $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$6 improvement over baseline methods (Liu et al., 2021).

These adaptation frameworks are important because they target one of the principal scalability constraints in drive-by inspection: the impracticality of obtaining labeled data from every bridge of interest. At the same time, both studies were evaluated on lab-scale bridges with controlled vehicles and fixed speeds, which makes domain adaptation a technical advance but not a complete solution to field generalization (Liu et al., 2020, Liu et al., 2021).

4. Vision-driven and digital-twin inspection pipelines

A different branch of drive-by bridge inspection uses visual and environmental sensing. A hybrid digital twin for bridge condition monitoring combines three near-real-time streams: YOLOv8 computer vision on an existing bridge-deck traffic camera, a Lighthill-Whitham-Richards traffic model, and public weather APIs. YOLOv8 produces detections

$$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$7

where $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$8 is the bounding box, $$F_{\text{res}(\omega)} = |F_{\text{front}(\omega)} - F_{\text{rear}(\omega)}|$$9 the vehicle class, $DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$0 the confidence score, and $DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$1 the number of detected vehicles. The LWR model propagates traffic density $DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$2 under

$DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$3

allowing detection of deceleration-driven shockwaves associated with repetitive loading and fatigue accumulation. Weather APIs provide temperature cycling, freeze-thaw activity, precipitation-related corrosion potential, and wind effects; Monte Carlo simulation propagates uncertainty; and Random Forest models map fused features to fatigue indicators and maintenance classification (Balijepalli et al., 14 Mar 2026).

UAV-based vision systems focus on surface detail and defect imagery. A benchmark of 23 models from YOLOv5, YOLOv6, YOLOv7, and YOLOv8 on COCO-Bridge-2021+ identified YOLOv8n, YOLOv7tiny, YOLOv6m, and YOLOv6m6 as balanced options for inference speed and accuracy. Reported $DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$4 values were 0.803, 0.837, 0.853, and 0.872, with inference times of 5.3 ms, 7.5 ms, 14.06 ms, and 39.33 ms, respectively; Jetson Nano throughput was 58.27 FPS, 36.31 FPS, 5.51 FPS, and 1.25 FPS (Phan et al., 2024).

Bridge-element segmentation from aerial inspection video has also been addressed through a semi-supervised self-training method. A Mask R-CNN with ResNet-50 backbone, temporal coherence analysis, and inspector-in-the-loop refinement achieved 91.8% precision, 93.6% recall, and 92.7% F1-score using 66 manually annotated images and 3.58 hours of inspector labeling time. The iterative update rule was

$DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$5

The paper also reported moderate cross-bridge generalization without adaptation, with F1-scores of 74.8% and 60.6% on two other bridges, reinforcing the importance of domain-specific adaptation (Karim et al., 2021).

For crack localization in reinforced concrete imagery, a cost-sensitive semantic-segmentation framework based on a SegNet-style encoder-decoder examined UW-MAP, MFW-MAP, and UW-ML strategies under severe crack/background imbalance. The weighted cross-entropy loss is written as

$DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$6

and the paper concluded that UW-ML provided the best F1-score and Mean Precision Accuracy trade-off on real-world crack images (Sajedi et al., 2019).

5. Autonomous UAVs, ground robots, and climbing robots

Autonomous path planning is a major subproblem in drive-by-style visual inspection. GATSBI addresses UAV surface inspection when the bridge geometry is not known beforehand. It creates a 3D occupancy map online from LiDAR, semantically segments bridge voxels, clusters candidate viewpoints into a Generalized Traveling Salesperson Problem, and replans in a receding-horizon loop. The objective is to find a minimum-length tour that visits at least one viewpoint in each cluster. In AirSim evaluations on five bridge models, GATSBI inspected 100% of all reachable bridge surface voxels on all bridges, whereas the frontier baseline inspected only 5–10%; in a hardware experiment, all 15 inspectable bridge voxels were inspected (Dhami et al., 2020).

For box girder bridges under GPS- and optical-flow-denied conditions, another UAV system combines local navigation routines, a supervisor, and a GTSP-based planner. The planner decomposes the bridge into planar polygonal surfaces and uses edge costs such as

$DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$7

while local navigation relies on two 2D Lidars, Hough-transform line extraction, and PID control to maintain offset from structural surfaces (Yu et al., 2019).

Ground and contact-based robotic systems extend drive-by inspection into NDE-rich regimes. One autonomous bridge deck robot integrates Impact-Echo, Ultrasonic Surface Waves, Electrical Resistivity, Ground Penetrating Radar, and high-resolution cameras, with EKF localization from dual RTK GPS, laser scanners, and IMU. It was reported as successfully deployed on more than 40 bridges in over ten U.S. states and capable of covering bridge decks up to 20 ft $DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$8 200 ft in 40 minutes (La et al., 2017). A related robotic system equipped with GPR, ER, and a camera used CLAHE, HOG features, and a Naive Bayes classifier for automated rebar picking, with reported rebar detection accuracy of 96.69% and precision of 99.59% on the Pleasant Valley Bridge, compared against RADAN7 (Le et al., 2017).

Steel-bridge inspection introduces adhesion and climbing constraints. A climbing robot with video cameras and a time-of-flight 3D camera performed on-the-move steel surface imaging, image stitching, ICP-based 3D registration, and automatic crack detection, reporting 93.1% correct detection for cracks greater than 3 mm width on 231 field images (La, 2017). A more recent intelligent magnetic inspection robot for ferromagnetic infrastructure used magnetic wheels and MobileNetV2, achieving 85% precision across six defect types and operating on vertical, inclined, and internal-corner steel surfaces (Tseng et al., 2024).

6. Operational realities, limitations, and emerging directions

A common misconception is that drive-by bridge inspection necessarily implies normal-traffic-speed, fully free-flow operation. The autonomous bridge deck robotic literature explicitly states that some platforms are not suited for at-speed inspection, that slow movement is likely required by current GPR and ER sampling methods, and that some degree of traffic control or partial lane closure may still be needed (Le et al., 2017). Even within vehicle-based indirect SHM, one study argues for an autonomous platform capable of maintaining a constant low speed precisely because accuracy and repeatability had not been achieved in previous studies (Hurtado et al., 1 Oct 2025).

Human oversight also remains central. A virtual reality-based training and assessment system for inspectors assisted by a drone, TASBID, integrates Unity simulation, a drone-control interface, real-time monitoring, and post-study assessment. In a pilot study with 22 participants, overall scores were mostly between 220 and 370 out of 400; efficiency was highest at about 92/100, safety lowest at about 69/100; and conformity and safety improved significantly after repeated training (Li et al., 2021). This suggests that even when sensing and navigation are automated, operator competence and human-drone collaboration remain part of deployment readiness.

Digital-twin deployment introduces organizational as well as technical constraints. In the Stava bridge case, a Digital Twin fed by IoT sensors, edge computing, and cloud analytics enabled online alerting, virtual inspection, and hybrid physics-plus-ML diagnostics. Indicators such as

$DI = \frac{1}{n} \sum_{j=1}^n (x_j - \hat{x}_j)^2,$9

were trended, anomaly rates accelerated sharply before closure, and offline analysis localized the problem to the southeast support/abutment. The case is presented as demonstrating the value of online monitoring and Digital Twins, while also emphasizing alert management, explainability, organizational readiness, and integration with risk- and condition-based maintenance (Hagen et al., 2024).

Current research directions are therefore heterogeneous rather than singular. They include purpose-built inspection vehicles with optimized mass-stiffness properties (Hurtado et al., 3 Oct 2025), scalable hybrid digital twins using existing traffic cameras and weather feeds (Balijepalli et al., 14 Mar 2026), domain-adaptive diagnosis without target labels (Liu et al., 2021), and human-in-the-loop visual pipelines for varied bridges in the National Bridge Inventory (Karim et al., 2021). Taken together, these developments indicate that drive-by bridge inspection is evolving from a single indirect-sensing technique into a broader monitoring paradigm spanning physics-based inference, computer vision, autonomous robotics, uncertainty quantification, and digital-twin-enabled maintenance decision support.