In modern infrastructure management, the health monitoring of bridge cables is critical for ensuring structural safety and longevity. Traditional contact-based sensors, while accurate, require extensive deployment and maintenance, often disrupting normal operations. With advancements in computer vision and unmanned aerial vehicle (UAV) technology, non-contact vibration measurement has emerged as a promising alternative. UAV drones offer mobility and flexibility, overcoming limitations of fixed cameras such as restricted field of view and resolution. However, challenges persist, including motion interference from UAV drones themselves and difficulties in target tracking in dynamic environments. This study addresses these issues by proposing a novel, target-free method for bridge cable vibration identification using UAV video. The approach integrates line segment detection, robust feature point tracking, and motion compensation to extract accurate vibration characteristics. Validation on real-world bridge inspection videos demonstrates high precision in frequency identification, paving the way for intelligent bridge monitoring systems.
The use of UAV drones for structural health monitoring has gained traction due to their ability to capture high-resolution video from various angles. In bridge applications, cables are key load-bearing components susceptible to vibrations from wind, traffic, and environmental factors. Accurately measuring these vibrations is essential for assessing cable tension and detecting potential damage. Vision-based methods leverage cameras to track displacements without physical contact, but when using UAV drones, the platform’s own motion introduces noise that must be separated from the structural vibrations. Previous studies have relied on fixed cameras with artificial targets or inertial measurement units (IMUs) for compensation, but these approaches lack generalizability or introduce cumulative errors. Our method eliminates the need for targets by exploiting geometric features of cables and background structures, enabling robust vibration extraction even in complex scenarios. By focusing on UAV drones as mobile platforms, we enhance the scalability and efficiency of bridge inspections, contributing to the development of low-altitude economy in transportation.
The core innovation lies in a framework that processes UAV video to isolate cable vibrations. As shown in the methodology overview, the process begins with video stabilization to mitigate jitter from UAV drones. Then, regions of interest (ROIs) for both cables and static background are selected, and linear features are detected using optimized line segment algorithms. Feature points are extracted from these segments and tracked across frames, with a filtering mechanism to ensure stability. The motion trajectories from background ROIs approximate the global movement of UAV drones, which is subtracted from the cable ROI motions to obtain pure vibration displacements. Post-processing techniques, such as detrending and high-pass filtering, remove residual noise, allowing for vibration analysis through frequency domain methods. This holistic approach ensures accuracy without relying on external sensors or manual markers, making it ideal for large-scale bridge monitoring with UAV drones.
To formalize the method, let the total displacement of a cable point in the video frame be denoted as $D_{total}(t)$, which consists of the cable’s true vibration $V_{cable}(t)$ and the UAV drone’s motion $M_{UAV}(t)$. We express this as:
$$ D_{total}(t) = V_{cable}(t) + M_{UAV}(t) $$
By tracking static background points, we estimate $M_{UAV}(t)$ as the average displacement of background features, represented by:
$$ M_{UAV}(t) = \frac{1}{N} \sum_{i=1}^{N} B_i(t) $$
where $B_i(t)$ is the displacement of the $i$-th background feature point, and $N$ is the number of stable points. The compensated cable vibration is then:
$$ V_{cable}(t) = D_{total}(t) – M_{UAV}(t) $$
This compensation is crucial for UAV drones, as their motion often overlaps with low-frequency cable vibrations, complicating direct frequency filtering. Our approach effectively decouples these components in the displacement domain, enhancing reliability for real-world applications.
The pre-processing stage involves frame stabilization to address jitter from UAV drones during flight. Given a video sequence with frames $F_k$ for $k = 1, 2, \dots, K$, we use Shi-Tomasi corner detection to extract feature points in the initial frame. Optical flow is applied to track these points across consecutive frames, generating affine transformation matrices $A_k$ that describe inter-frame motions. The cumulative displacement $C_k$ relative to the first frame is computed by summing these transformations, and smoothing is applied to $C_k$ to obtain stabilized frames $F’_k$. This step reduces errors from UAV drone vibrations, ensuring that subsequent analyses focus on structural movements. The stabilization algorithm’s effectiveness is quantified by the reduction in high-frequency noise, which we assess through power spectral density analysis in later stages.
For feature extraction, we utilize the Line Segment Detector (LSD) algorithm to identify linear contours in both cable and background ROIs. UAV drones capture videos where cables appear as prominent edges, and background structures like bridge towers exhibit similar features. The LSD algorithm detects line segments based on gradient regions, but raw outputs may be fragmented due to environmental factors. We implement a post-processing pipeline to merge broken segments: two segments are combined if their angle difference $\Delta \theta$ and endpoint distance $\Delta d$ satisfy thresholds, such as $\Delta \theta < 15^\circ$ and $\Delta d < 10$ pixels. Additionally, we connect adjacent segments with continuity conditions, e.g., distance less than 15 pixels and angle difference under $5^\circ$. From the optimized lines, feature points are sampled at uniform intervals, creating a set for tracking. This method leverages the inherent geometry of scenes captured by UAV drones, avoiding the need for artificial targets.
The tracking process employs optical flow to monitor feature points across frames. For each ROI, we assign unique IDs to points in the first frame and verify their continuity throughout the sequence. Points that persist across all frames are deemed stable and used for displacement calculation. This filtering mitigates issues like occlusion or lighting changes common in UAV drone videos. The stability criterion is defined as a binary function $S_i$ for point $i$, where $S_i = 1$ if tracked in all frames, else $S_i = 0$. The set of stable points $P_{stable}$ is then:
$$ P_{stable} = \{ p_i | S_i = 1 \} $$
We track both cable and background points simultaneously, enabling real-time motion estimation. The displacement for each point is computed as the Euclidean distance from its reference position, averaged over $P_{stable}$ to represent ROI motion. This approach ensures robustness, as UAV drones may experience sudden movements that could otherwise corrupt data.
Motion compensation is pivotal for isolating cable vibrations. Using background ROI trajectories, we approximate the UAV drone’s global motion as a time series $M_{UAV}(t)$. For each frame, the cable ROI displacement $D_{cable}(t)$ is adjusted by subtracting $M_{UAV}(t)$, yielding the pure vibration signal $V_{cable}(t)$. In practice, we calculate $M_{UAV}(t)$ as the mean displacement of all stable background points, as shown earlier. This step effectively removes low-frequency components introduced by UAV drones, which is essential for accurate frequency analysis. The compensated signal may still contain linear trends from residual errors, so we apply detrending by fitting a line $L(t) = at + b$ to $V_{cable}(t)$ and subtracting it:
$$ V’_{cable}(t) = V_{cable}(t) – (at + b) $$
where $a$ and $b$ are estimated via least squares. A high-pass filter with cutoff frequency $f_c$ (e.g., 0.1 Hz) is then used to eliminate low-frequency noise, producing the final vibration signal for analysis.
Vibration characteristics are extracted using frequency domain methods. The processed signal $V’_{cable}(t)$ is transformed via Fast Fourier Transform (FFT) to obtain the power spectral density $P(f)$:
$$ P(f) = \left| \int_{-\infty}^{\infty} V’_{cable}(t) e^{-i2\pi ft} dt \right|^2 $$
Peaks in $P(f)$ correspond to natural frequencies of the cable. For multi-cable systems, we partition the ROI into subregions and compute frequencies separately, taking the median as the representative value. The fundamental frequency $f_0$ is identified as the lowest significant peak, and higher-order modes $f_n$ for $n = 1, 2, \dots$ are also detected. Cable tension $T$ can be estimated using string vibration theory, where for a cable with mass per unit length $\mu$ and length $L$, the $n$-th frequency relates to tension as:
$$ f_n = \frac{n}{2L} \sqrt{\frac{T}{\mu}} $$
This formula assumes ideal conditions; in practice, we use multiple modes to improve accuracy. The integration of UAV drones facilitates data collection from various angles, enhancing mode shape identification for tension estimation.
To summarize the methodology, Table 1 outlines the key steps and their purposes in the context of UAV drone-based monitoring.
| Step | Description | Role in UAV Drone Application |
|---|---|---|
| Video Stabilization | Reduce inter-frame jitter from UAV motion | Minimizes noise induced by UAV drones during flight |
| ROI Selection | Define cable and background regions | Leverages UAV drone flexibility to capture relevant areas |
| Line Segment Detection | Identify linear features using LSD algorithm | Exploits geometric patterns visible in UAV drone footage |
| Feature Point Tracking | Monitor points via optical flow with stability filtering | Addresses challenges from UAV drone movement and environmental changes |
| Motion Compensation | Subtract background motion to isolate vibrations | Decouples UAV drone motion from structural responses |
| Signal Post-processing | Detrend and filter to remove noise | Enhances signal quality for accurate analysis with UAV drones |
| Frequency Analysis | Extract modes via FFT and compute cable tension | Provides structural health indicators using UAV drone data |
The experimental validation was conducted on the Jiangyin Yangtze River Bridge, a long-span suspension bridge with multiple cables. We deployed a DJI Matrice 350 RTK UAV drone equipped with a Zenmuse H30T camera, capturing videos at 30 Hz resolution of 3840 × 2160 pixels. The UAV drone was flown to hover near selected cables, recording 90-second clips for each. This setup mimics real inspection scenarios where UAV drones are used for routine monitoring. The cables, designated as 82N, 82S, 83N, 83S, 84N, and 84S, were analyzed using our method. Prior contact sensor measurements provided baseline fundamental frequencies for comparison, ensuring objective evaluation of accuracy.
In pre-processing, the stabilization algorithm significantly reduced frame jitter, as evidenced by the smoothed displacement time series. For example, cable 82N showed a noise reduction of approximately 40% in high-frequency components after stabilization, confirming the effectiveness for UAV drone videos. The line detection and optimization successfully extracted cable edges, with an average of 15-20 stable feature points per ROI. Tracking maintained continuity for over 95% of points across frames, demonstrating robustness against disturbances common in UAV drone operations.
The motion compensation step yielded clean vibration signals. Figure 1 illustrates the displacement before and after compensation for cable 82N: the raw data included low-frequency drifts from the UAV drone, while the compensated signal exhibited clear oscillatory patterns. The power spectral density revealed distinct peaks corresponding to natural frequencies. Table 2 compares the fundamental frequencies obtained from our UAV drone-based method with contact sensor measurements, highlighting the precision achieved.
| Cable ID | Sensor Measurement (Hz) | UAV Drone Method (Hz) | Relative Error (%) |
|---|---|---|---|
| 84N | 0.836 | 0.805 | 3.71 |
| 84S | 0.815 | 0.805 | 1.23 |
| 83N | 0.852 | 0.834 | 2.11 |
| 83S | 0.848 | 0.834 | 1.65 |
| 82N | 0.895 | 0.878 | 1.90 |
| 82S | 0.914 | 0.878 | 3.94 |
All errors are below 5%, meeting engineering standards for cable tension estimation. Moreover, the method extracted multiple vibration modes, as shown in the spectrum for cable 82N, where peaks at 0.878 Hz, 1.756 Hz, and 2.634 Hz correspond to the first, second, and third harmonics, respectively. This multi-mode capability enhances tension calculation reliability. The use of UAV drones enabled efficient data collection from multiple cables without physical access, reducing inspection time by over 50% compared to traditional methods.
The results underscore the viability of UAV drones for structural health monitoring. The compensation algorithm effectively separated UAV drone motion from cable vibrations, even when frequencies overlapped in the 0.3-1.5 Hz range. The linear trend removal and high-pass filtering further refined signals, with the cutoff frequency optimized at 0.1 Hz based on empirical tests. The accuracy is attributed to the synergistic use of background tracking and geometric feature extraction, which are uniquely facilitated by the mobility of UAV drones. In field tests, all cables were analyzed successfully, with no instances of tracking failure or compensation errors, validating the method’s robustness in real-world conditions.
For broader application, the method’s computational efficiency is crucial. On a standard laptop, processing a 90-second UAV drone video takes approximately 5 minutes, including stabilization, tracking, and analysis. This allows for near-real-time monitoring during inspections with UAV drones. The algorithm’s parameters, such as LSD thresholds and filter cutoffs, can be tuned for different environments, making it adaptable to various bridge types and UAV drone models. As UAV drones become more prevalent in infrastructure management, such automated methods will drive advancements in intelligent transportation systems.
In conclusion, this study presents a comprehensive framework for bridge cable vibration identification using UAV video. By integrating computer vision techniques with motion compensation, we achieve target-free, high-precision measurement of vibration characteristics. The method addresses key challenges posed by UAV drones, such as ego-motion interference and tracking instability, through innovative algorithms for line detection, feature point management, and signal processing. Validation on a major suspension bridge confirms fundamental frequency errors under 5%, with successful extraction of multiple modes for tension estimation. The approach leverages the flexibility of UAV drones to enable scalable, non-contact inspections, reducing costs and improving safety. Future work will focus on enhancing automation via deep learning for multi-cable recognition and extending applications to other structures, further solidifying the role of UAV drones in smart infrastructure monitoring.
The implications for the low-altitude economy are significant, as UAV drones offer a cost-effective solution for large-scale bridge assessments. By eliminating the need for sensors or targets, our method lowers barriers to adoption, allowing routine inspections with minimal disruption. As UAV drone technology evolves, integrating higher-resolution cameras and advanced stabilization will further boost accuracy. We envision a future where fleets of UAV drones autonomously monitor infrastructure, providing real-time health data to engineers. This study contributes a foundational step toward that vision, demonstrating that with robust algorithms, UAV drones can revolutionize structural health monitoring in the transportation sector.