01978nas a2200241   4500000000100000000000100001008004100002260001200043653002000055653002100075653002000096653001900116653001700135100003000152700003800182700003000220245006700250856005800317300000900375490001300384520132500397022001401722       9998                            d  c08/202410aComputer vision10aMachine Learning10aNeural Networks10aRailway Safety10aSurveillance1 aDomingo Martínez Núñez1 aFernando Carlos López Hernández1 aJ. Javier Rainer Granados00aAutomatic Surveillance of People and Objects on Railway Tracks  uhttps://www.ijimai.org/journal/bibcite/reference/3478  a1-100 vIn press3 aThis paper describes the development and evaluation of a surveillance system for the detection of people and objects on railroad tracks in real time. Firstly, the paper evaluates several background subtraction techniques including CNNs and the object detection library called YOLO. Then we describe a novel strategy to mitigate the occlusion caused by the perspective of the camera and the integration of an alarms and pre-alarms policy. To evaluate its performance, we have implemented and automated the control and notification aspects of the surveillance system using computer vision techniques. This setup, running on a standard PC, achieves an average frame rate of 15 FPS and a latency of 0.54 seconds per frame, meeting real-time expectations in terms of both false alarms and precision in operational mode. The results from experiments conducted with a publicly available recorded video dataset from Metro de Madrid facilities demonstrate significant improvements over current state-of the-art solutions. These improvements include better accident anticipation and enhanced information provided to the operator using a standard low-cost camera. Consequently, we conclude that the approach described in this paper is both effective and a more practical, cost-efficient alternative to the other solutions reviewed.  a1989-1660