A multi-scale image and dynamic candidate region-based automatic detection of foreign targets intruding the railway perimeter

Abstract This paper focuses on image characteristics of railway monitoring scenes, and proposed a multi-scale image and dynamic candidate region-based automatic detection of foreign targets intruding the railway perimeter. The algorithm first constructs a multi-scale image set for each frame, then obtains a dynamic candidate region of the intrusion target according to the law of gray projection curve variation for the low-resolution image. The foreground image, containing the intrusion target, is obtained using the background differential method for the high-resolution image, then the intruding target is identified before fusion of the results from the two methods. Three railway surveillance videos are used to validate the algorithm comparing with other methods, currently used in the railway scene. Overall, the proposed algorithm guarantees high-accuracy real-time detection of intruding targets, while requiring little computational resources, which is able to achieve 89.3% Precision and 0.028 s/frame with CPU. This technology can reduce the workload of manual inspections, and improve the detection accuracy through the improvement of the model can also reduce the manual review work caused by false alarms.