Subway Vehicle Anomaly Detection Method Considering Comparison of Image Multi-feature Discrepancy Fusion

Objective A subway vehicle anomaly detection method based on fusion comparison of images multi-feature discrepancy is proposed, aiming to address the issues of high workload and low efficiency in current manual inspection of subway vehicles.

Method First, the train pass-by exterior image is used as the detection image, and the latest historical image of the same train with the same name is obtained from the database as the template image. Then, a 64 px×64 px slider is used to simultaneously scan both the detection and the template images. Within each slider region, the color difference, the HOG (histogram of oriented gradients) difference, and the edge matching difference between the two images are calculated respectively. Finally, these features are fused using different pre-set feature weights to compute a difference score, which is used to determine if an anomaly exists within the slider region. The effectiveness of this detection is validated through actual tests conducted at Beijing subway vehicle inspection points.

Result & Conclusion  The subway vehicle anomaly detection method based on image multi-feature discrepancy fusion comparison can accurately identify simulated faults on the vehicle's bottom. So far, it has successfully detected defects such as foreign objects, scratches, oil leaks, missing and loose of screws, achieving an accuracy rate of over 98%.