Research On The Theory Of Visual Object Tracking And Its Application Based On Particle Filter

Object tracking through image sequences has been an active field in computer vision, image processing and patern recognition. It has a wide range of applications in military guidance, visual surveillance, visual navigation of robots, human-computer interaction, and medical diagnose, etc. Tracking is caried out by analyzing the image sequence from the sensor, identifying the independently moving regions or those that users are interested in, and locating them in the image. It's of great importance not only because it provides the track of the object, but also it offers reliable information for object recognition, scene analysis and other applications. As a topic with many applications, object tracking has drawn much attention of researchers. Many institutes have done a lot of work on it and got achievements. However, visual target tracking often becomes very difficult due to complex image backgrounds and the target motion. There are still many problems to solve in order to build a robust and practical tracking system. In this dissertation, the research is focused on the tracking problems of single target and multiple targets.The main research work in the dissertation is as follows:1. No feature-based vision system can work unless good features can be identified and tracked from frame to frame. We extend the well-known Shi-Tomasi-Kanade tracker to an efficient method of object tracking with motion prediction. Then object tracking are implemented by matching local invariant features which are learned online. Interest points extracted with the proposed Harris-SIFT detector can be adapted to affine transformations and give repeatable results. The experimental results show that the proposed method is capable of tracking objects under partial or severe occlusions. Then the algorithm and applications related to particle filter are surveyed. The problems of particle filter are discussed and some improvement methods are illustrated. Finally, further research directions are pointed out.2. To tackle the divergence of classical particle filter method for multiple objects tracking in image sequences, a new particle filter, called pseudo-particle filter (PPF), is proposed. The PPF invokes subset particles of generic particle filters to form a continuous estimate of the posterior density function of the objects. After sampling-importance resampling (SIR), the subset particles converge to the observations. It is proved that, using appropriate kernel function of mean-shift algorithm, we can get the subset particles of the observations and the fixed points of clustering results as the state of the objects. A multiple objects data association and state estimation technique is proposed to resolve the subset particles correspondence ambiguities that arise when multiple objects are present. Experimental results demonstrate the efficiency and effectiveness of the algorithm for single and multiple objects tracking.3. A potential weakness of generic particle filters is that the particle-based approximation of filtered density is not sufficient to characterize the tail behavior of true density, due to the use of finite mixture approximation; To alleviate this problem, we propose a general hierarchical particle filtering framework for designing an optimal proposal distribution. The essential idea is to augment a second filter's estimate into the proposal distribution design. We shall see that several existing improved particle filters can be unified into our general framework. Based on this framework we further propose novel variant algorithms for robust and efficient visual tracking.4. The nonsubsampled Contourlet (NSCT) not only overcomes the disadvantage of wavelet, the nonoptimal basis for one dimensional singularity, but also improves the edge preservation for the shift-invariance. An image fusion algorithm based on directional windows statistics in NSCT domain is developed. Source images are firstly decomposed to the domain of the nonsubsampled contourlet transform, the image fusion is then implemented in subbands with different scale and direction combining with region statistics. Regional variance and local energy are adopted as fusion rules in Iowpass and highpass subbands, respectively. Finally the fused image is obtained through inverse transform. Experimental results show that the directional windows in NSCT can detect image features more effectively and the fused image has better subjective visual effect. Compared to traditional gradient pyramid algorithm and wavelet based algorithm, the fusion quality index by the proposed algorithm is outperformed the formers.5. In order to satisfy the high precision and rapidness requirement of the romote tracking systems of objects, a tracking system based on DSP is implemented. the coding architecture of H.264 Baseline Profile is introduced first, and some reasons for limiting speed are presented according to the analysis of the coding architecture. Besides, several implementation techniques for optimization are presented. And one kind of real-time voice-band system based on DSP and ASIC is proposed, it fetches the instruction which the host system sends out, gives the response in real-time. Novel algorithms of FAT32 file system and AVI file format for increasing the compatibility of video storage based on DSP is given. Based on analysis of the video, automatic command system determines the status of flying parameter. It sends out flying assistant commands in real-time.