PEDESTRIANS TRACKING AT SIGNALIZED INTERSECTIONS USING MULTI-CAMERAS AND POSE ESTIMATION ALGORITHM

Pedestrians’ detection is considered one of the key problems that faces traffic safety researchers in recent years. The different patterns of motion, various dressing colors, partial occlusions, and pedestrians’ positions with respect to the detecting devices pose additional challenges for the monitoring process. Various approaches were employed to detect pedestrians including surveillance cameras, LiDARs, Infrared Light Emitting Diodes (IR LED) stereo cameras, and microwaves energy beams. These techniques have seen extensive application in the realm of pedestrians’ detection and tracking, resulting in the establishment of robust safety assessment methodologies. However, despite their effectiveness, the widespread deployment of these techniques is hindered by several constraints. To address these challenges and provide an optimized detection framework, this study introduces an approach that combines surveillance cameras with a distinct Convolutional Neural Network (CNN)-based pose estimation algorithm to detect and track pedestrians at signalized intersections, resulting in accurate and reliable outputs. A geometrical spatial proximity method built on linear and curvilinear perspectives principles is proposed to restore the pedestrian joints’ coordinates from image plane to top-down view. Consequently, these coordinates are further clustered, and integrated to construct the pedestrians’ trajectories. The presented approach provides an efficient and accurate assessment tool for pedestrians’ motions at signalized intersections. By leveraging the strengths of the integration between surveillance cameras and CNN-based pose estimation algorithm, this research aims to contribute significantly to the advancement of traffic safety practices.