Pedestrian Multiple Object Tracking  Using Deep Learning

Abstract: In this thesis, the aim is to examine the viability of Deep Neural  Network (DNN) based Multi-Object Tracking approaches for tracking  pedestrians. The tracking results are used for Autonomous Driver Assistance System (ADAS). The process of tracking multiple agents  across video is termed as Multiple Object Tracking (MOT). Using only  pedestrian detection for ADAS is inferior to using pedestrian tracking  to determine the next action to be taken. By tracking pedestrians, one  can predict their next position with great accuracy. Furthermore,  pedestrians may not be detected in certain instances such as occlusion, and thus the system fails to deliver the stability needed  for safety priorities. For detecting the pedestrians, a custom version  of YOLO v3 is applied. The most common problems that arise in tracking  are ID switches and losing track of objects due to occlusion. Two  tracking methods Deep SORT and SORT-OH are utilised to track  pedestrians. The Deep SORT approach uses motion feature obtained by  Kalman filter and appearance feature obtained using CNN feature  encoder developed for Person Re-ID. SORT-OH is an approach that  handles occlusion and re-identifies targets using geometric cues. By  the means of ablation study, it is found that Re-Identification plays  a critical role in obtaining good tracking results. Also by using  different object detectors DPM, SDP and Faster R-CNN, it could be  concluded that the quality of detection plays a vital role in the  performance of the Object Tracking approach. Both Deep SORT with DNN based Re-ID and SORT-OH with geometric cue-based Re-ID features  perform similarly, demonstrating the potential of DNN techniques in  object tracking. Both the MOT methods obtained good results by  minimizing ID switches and handling occlusion, even in a difficult  situation with blur and varying illumination.