Urban road users detection and velocity estimation from top-view fish-eye imagery under low light conditions

In this study we present a technique which performs pedestrians and vehicles detection through the use of off-the-shelf object detectors alongside existing optical flow based velocity estimation techniques. Road users detection and apparent movement quantification are carried out respectively by YOLOv4 and FlowNet2.0. The speed of the users is then estimated from the average optical flow in each bounding box. Experimental results show that our proposed methods can effectively detect road users and estimate their velocity and movement direction under low light and low contrast urban video scenes. The videos are made with a low cost fish-eye RGB camera placed in a vertical position at a height of less than 10 m. The estimated speed by taking into account the deformations generated by the wide-angle lens, is in very good agreement with reality.