Multi-target visual tracking with a robust gaussian mixture probability hypothesis density filter

Abstract

Multi-target visual tracking (MTVT) has emerged as an active research topic in the past two decades because of its widespread applications in many areas, including intelligent surveillance, smart rooms, visual human–computer interfaces, autonomous robotics, augmented reality, and video compression. However, MTVT remains a challenge because of several factors, including noisy video data, varying number of targets, mutual occlusion, and data association problems. The Probability Hypothesis Density (PHD) filter, which propagates only the firstorder statistical moment of the full target posterior, has been shown to be a computationally efficient method that can avoid the challenge brought by data association for multi-target tracking. The Gaussian mixture PHD (GM-PHD) filter, whose posterior intensity is estimated by a sum of weighted Gaussian components that can be propagated analytically in time, is a closed-form solution to the PHD filter recursion. In this thesis, an MTVT system that combines object detection with a robust GM-PHD filter is developed to track multiple moving targets. To accurately track newborn targets from noisy data, an effective birth intensity estimation method based on entropy distribution and coverage rate is proposed. Birth intensity is first initialized using previously obtained target states and measurements. The measurements are obtained by object detection and are classified into birth measurements and survival measurements. The initialized birth intensity is then updated using currently obtained birth measurements. In the update stage, the entropy distribution is first applied to remove the noises within the birth intensity that are irrelevant to the birth measurements. Then, the coverage rate between each birth intensity component and the corresponding birth measurement is computed to further eliminate the noises. With this effective birth intensity estimation method, the proposed MTVT system can accurately track the newborn targets by reducing the interferences by the noises. To track multiple targets that move close together, an effective weight penalization method incorporated with a spatial color appearance model is proposed. The weight matrix of the updated weights between the measurements and the predicted target states is constructed. A criterion is proposed to search and determine the ambiguous weights in the constructed matrix. A spatial color-based appearance is modeled and used to penalize the ambiguous weights. The weights after penalization are normalized to form a new weight matrix. With this new weight matrix, the GM-PHD filter can accurately update the target states and thus achieve accurate tracking results. To track targets in mutual occlusion, a game-theoretical mutual occlusion handling method incorporated with an improved appearance model is proposed. A simple two-step occlusion reasoning method is proposed to determine the occlusion region. An improved spatial color-based appearance with interferences by other targets within the occlusion region is modeled and used to measure the similarity between the candidate and the target model. Compared with the conventional color histogram-based appearance model, the improved model is more robust even when the targets involved in the occlusion region have similar color distributions. An n-person, non-zero-sum, non-cooperative game is constructed to bridge the joint measurements estimation and the Nash equilibrium of the game. The individual measurements originating from the targets within the occlusion region are regarded as the players in the constructed game competing for maximum utilities using certain strategies. The Nash equilibrium of the game is the optimal estimations of the locations of the players within the occlusion region, which are equivalent to the optimal segmentations of the measurement originating from the targets in mutual occlusion. Experiments conducted on several publicly available data sets demonstrate the good performance of the proposed MTVT system.

Date of Award	2 Oct 2013
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	You Fu LI (Supervisor)