Real-Time Cyclist Tracking in a Video using CNN and Deep Sort
Abstract - Due to ever increasing environmental safety concerns, cycles and electric vehicles promise to be the most popular modes of transport. Further, the advancements in autonomous electric vehicles pose a greater threat for pedestrians and cyclists, in terms of road safety, who would be vulnerable as road users. Thus, cyclist detection and tracking in videos is a new dimension in multi-object tracking problems encountered in computer vision research. Calculating speed of cyclist is also important for autonomous vehicles to avoid collision between cyclist and vehicle. In this paper, a novel approach for detection, tracking and calculating the speed of cyclist in videos is proposed. The cyclist detection is done by modified YOLOv3 and real-time cyclist tracking is done by employing Deep SORT. The movement representation and data association algorithms are used for cyclist tracking and the optical flow is used to calculate the speed of cyclist. The proposed algorithm is experimented on two benchmark datasets, namely, KITTI and SCD datasets, of videos of real scenes with cyclists.
Keywords - Convolutional Neural Network, Video Processing, Multi Object Tracking, Data Association