On an evaluation of tracking performance improvement by SMC-PHD filter with intensity image of pedestrians detection over on-board camera using neural network

Abstract

Performance evaluation of multiple pedestrian tracking with/without the particle filter technique has been conducted by proposing some elaborated criteria for evaluation in terms of 1) detection evaluation for each frame, and 2) tracking evaluation for each image sequence. We cope with non-triviality on performance evaluate of multiple pedestrians detection and tracking under the situation of having false positive and false negative, true positive and true negative, and swapping of tracking targets with respect to different pedestrians as well as among false tracks and true tracks. Evaluation results with comparison among A) Nearest Neighbour(NN), B) Plain mode of particle filter, and C) Weighted model of particle filter are summarized as follows. By truth detection rate criterion, A) NN is the worst performance, while B) Plain is better than C) Weighted. By Swapping ID criterion, performance is improved by B) Plain and C) Weighted with C) being slightly better than B). However, by the criterion of termination of tracking, B) and C) are not necessarily better than A), rather worse than A), and B) Plain is worse than C) Weighted. This means that short term tracking performance has been improved by particle filter. Also, an elaboration in C) Weighted to consider the change of target size improve the performance than B) Plain.