Video-Based Automated Pedestrians Counting Algorithms for Smart Cities

Abstract

Smart cities possess several technologies in collecting pedestrian activity data, which may be used to manage city planning. A growing body of research exists on video processing based pedestrian counting methods, due to the development of new computer vision techniques. This research reviews different, vision-based methods for counting pedestrians and applies a specific counting method which is formed by a combination of You Only Look Once Version 3 (YOLOv3) and Simple Online Real-time Tracking (SORT) with a deep association metric. The results suggest that although clustering, as well as the direction and intensity of pedestrian traffic, achieves a minimal effect on the count, occlusion constitutes the main source of errors. Adequate training may serve to increase accuracy.