Study of Identifying Jaywalkers by Analyzing Camera Data

Yifan Li, Purdue University

Abstract

Jaywalking is an unlawful pedestrian behavior. In the United States, jaywalking leads to hundreds of pedestrian fatalities each year. Existing studies analyzed pedestrian behaviors by computer vision platforms. However, many of those existing platforms were not capable of detecting the single pedestrian in crowded traffic scenes. Those platforms were implemented based on the motion-based pedestrian detector, and the motion-based detector cannot detect pedestrians who are partially occluded or close to other moving objects. Furthermore, those existing platforms did not focus on jaywalker identification; those platforms need manual annotation to distinguish jaywalkers among all pedestrians. This study proposes a jaywalker-identifying metric and a jaywalker monitor platform. The proposed metric can identify jaywalkers using traffic light information; thus the platform can identify jaywalkers autonomously based on the proposed metric. The proposed platform implements an image-based pedestrian detector using You Only Look Once (YOLO) neural network. This YOLO-based detector does not need pedestrian's motion information to detect pedestrians; therefore, the proposed platform can discover single jaywalker in crowded traffic scenes. The proposed jaywalker monitoring platform is evaluated on a pedestrian video dataset. This study compares the performance of a motion-based pedestrian detector and a YOLO-based pedestrian detectors. The study also explores the performance potential of the YOLO-based detector, and the experiment results show that adjusting the input size of YOLO neural network can improve recall and precision rates for detecting pedestrians.

Degree

M.S.E.C.E.

Advisors

Lu, Purdue University.

Subject Area

Computer Engineering|Electrical engineering

COinS