Object Tracking Based On Multi-feature Fusion And Two-Stage Particle Filter

Visual object tracking is designed to imitate the motion sensibility of human vision. It is the basis of video ananlysis and understanding. Visual object tracking has a wide range of applications in intelligent video surveillance, intelligence transportation, human-computer interaction, and military guidance, etc. Robust and accurate tracking results are hard to obtain when using tracking algorithm with single feature in complex environments. In order to solve this problem, tracking strategy which based on feature fusion has been put forward. T he product rule, weighted sum rule and layered fusion rule are the three common rules of feature fusion. T he product rule is simple and effective, but it has a poor anti-noise ability.T he weighted sum rule has a str onger anti-noise ability but a poorer fusion effect. T he disadvantages of layered fusion rule are the uncertainty of priori knowledge and the fixed using order of features. In addition, multiple feature fusion tracking strategy greatly increases the amou nt of calculation, which make the real-time visual tracking is difficult to improve.To solve those problems, a tracking algorithm that fuses multiple features based on a two-stage particle filtering framework is proposed. An improved adaptive particle filt er method is also proposed to improve the real-time performance.In this paper, tracking algorithm based on feature fusion strategy is studied in the particle filter framework. Features of color and texture are selected as the fusion feature. Research content including particle filter algorithm and its application in visual object tracking, multi-features fusion strategy and adaptive particle filter method. Specific contents as follows:1. The basic theory of particle filter algorithm is explained in deta il and the general steps of particle filter method is given. Target state model, state transition model and observation model are established at first, T hen initial particle set is obtained through random sampling from initial probability distribution of target state. particles are propagated through the state transition model(prediction) and the particle weights are updated according to the observation model, the posteriori probability density distribution of target state can be approximated by the weighted particle set. Finally the minimum mean squared error(MMSE) estimation or maximum a posteriori probability(MAP) estimation is used to obtain the final estimate of target state. T hrough the recursive prediction and weights updating process, the visual object tracking task will be completed.2. A tracking algorithm that fuses multiple features based on a two-st age particle filtering framework is proposed for the visual target tracking in complex environments. T he algorithm can be divided into two steps. In the first stage,a particle filter with single feature which has the best tracking performance is used to estimate an initial tracking result. With the initial estimation, another particle filter with feature fusion strategy based on weighted sum rule i s used to determine the final predicted state in the second stage. In the process of tracking, feature uncertainty measurement method is used to measure a feature's tracking performance and determine its weight in feature fusion strategy based on weighted sum rule. Comparative experiments are performed and the results show that the proposed algorithm is more accurate and robust than the single feature, the product rule, the weighted sum rule and the layered fusion rule tracking algorithms.3. To improve the real-time performance of multiple feature fusion target tracking method, an improved adaptive particle filter algorithm is proposed. In this method,the particle number and the noise variance of state transition model to are adjusted by measuring the redundancy of particle collection use KL divergence. In order to ensure the accuracy of observation likelihood function, a simplified EMD is introduced into the video target tracking. Finally, a comparative experiments are performed.T he results show that the improved algorithm can effectively and steady track the motion objects, with better real-time and robustness performance.