Research On Moving Target Tracking Method Based On SIFT And Particle Filter

Intelligent video surveillance technique is one of the computer vision hotspots, and many scholars are devoted to this research area; in the area of financial, public security, military and so on, it has great prospect. Intelligent video surveillance refer to using computer vision, image processing and machine learning and so forth knowledge to analysis video without human beings interference. The key technique to intelligent video surveillance is target tracking, which is also the basis of behavior and target identification, so the research of this theory and application is significant.Motion target tracking method can be divided into two categories:probability tracking method and Deterministic tracking method. Probability method has become the mainstream method for its stability and reliability, in which kalman filter and particle filter is the classic method. Kalman filter applicability is confined to a certain area, because it has strict limit for system model and posteriori distribution, and only can handle non-linear, non-gauss, and non-multiple model, particle filter, different from kalman filter, don’t require specific system model, can keep its state as multiple model distribution, and not easily affected by noise wave, so it is widely applied in tracking area; but normal particle filter tracking algorithm is a great challenge because of the complexity of actual tracking scene.This paper is based on color histogram particle filter algorithm, analyze some problems exist in this algorithm and suggest improvement method accordingly. particle filter algorithm based on color histogram only use target’s color characteristic to update particle weight, so when background color and target color distribute similarly or target get blocked, it is easy to lose target during tracking process. SIFT has relatively high stability, but only using SIFT is not enough for the description of small target. This paper handle problems mentioned above with particle filter taking SIFT characteristic and color characteristic into consideration. Experiences show that this algorithm can effectively improve tracking accuracy when target blocked or target color and background color mixed up.