Presenter: Dr. Mufleh Al-Shatnawi, York University
Topic: Improving Real-time Pedestrian Detection using Adaptive Confidence Thresholding and Inter-Frame Correlation
The pedestrian detection algorithms form a key component in the multiple pedestrian tracking (MPT) systems. Despite efforts to detect a pedestrian accurately, it is still a challenging task. We propose a novel and efficient online method to improve the performance of the multiple person/pedestrian detector by introducing novel post-processing steps. These steps use an adaptive approach to determine both area and confidence score constraints for the output of any given multiple pedestrian detector. In this paper, we focus on pedestrian detection in video surveillance applications that require an automated, accurate and precise pedestrian detection algorithm. We demonstrate that the new steps make the multiple pedestrian detector more accurate, precise and tolerant to false positive detections. This is illustrated by evaluating the performance of the proposed method in test video sequences taken from the Pedestrian Detection Challenge, Multiple Object Tracking Benchmark (MOT Challenge 2017).
© 2018 Compute Ontario