The Research Of Object Tracking Algorithm Based On Blocking Partition Particle Filter

Object tracking is the process of acquiring the target motion parameters according to the information about the moving object, which is acquired by the device of sensor, then processes and analyzes the information and gets the result according to the corresponding algorithm. The technology is widely used in monitoring, obstacle avoidance, navigation, automation control and so on. Aiming at the problem of low accuracy due to object deformation, the thesis proposed tracking algorithm based on blocking partition particle filter, and compared the new algorithm with Mean Shift and Camshift on object tracking by the method of stimulation.Firstly, this paper deeply analyzes the algorithm of Particle Filter. Particle Filter algorithm provides a theoretical framework, it needs the state and observation model, the state model is to provide predictive value for the state, and the predicted value is corrected by the observation model according to the observations. The technology of re-sampling and the choice of importance density function are important parts of the algorithm. There has been various algorithms through the improvement of the above two technologies, such as the Likelihood Particle Filter, Auxiliary Particle Filter and other methods. This paper also has analyzed the impact of state noise and observation noise on different algorithms.Secondly, this paper introduces several common methods of detection and tracking. The detection methods include temporal differencing, background subtraction, etc. Mean Shift and continuous adaptive mean shift(Camshift) algorithm are now commonly used in object tracking. Both of them are based on the color histogram. Mean Shift, uses the mean shift vector by iteration, so that the target can converge to the high density sample area. But the tracking window of mean shift is fixed, when the object deformation occurs or two adjacent frames overlap less, the effect of tracking gets worse; Camshift can adaptively adjust the size of the tracking window, but the tracking window defaults to the square, when the target becomes deformed seriously, the effect of tracking decreases in efficiency.Finally, an improved algorithm was proposed. The new algorithm used particle filter algorithm as its framework. After obtaining the size and position of the moving object, it transferred the RGB model into HSV model, and then constructed the color histogram of the target region according to H component, next built the Probability Distribution Image of the whole region based on the color histogram of target region. Next step, it equally divided the target region into N equal parts, each of these small region was considered as a "particle", and then determined the weight of each region according to the sum of the probability distribution of all pixel in each small region,and re-sampled the "particles" according to their weight, then calculated the mean value of the central value of each small region after resampling. This algorithm still had high stability and reliability in the conditions of changes of background, the object deformed and other harsh conditions.