Researsh On The Head-shoulder Video Detection And Tracking Algorirhms For DSP

Pedestrian detection and tracking is the field of intelligent video analysis is an importantresearch topic. Because the body’s head and shoulders showing relatively fixed similar "Ω" shapeand can not easily be blocked, etc., for pedestrians to accurately detect and track provides aprerequisite. Aiming at the scene of a single fixed camera, motion detection and tracking of humantargets were studied. The main work is as follows:First, the moving object extraction areas, summarizes the current moving target detection ofseveral commonly used methods. Against the background of the video surveillance system fixedthis feature, using background subtraction to extract moving targets. After background subtraction,the first use normalized color information to eliminate the shadow of moving target, and then usemathematical morphology methods to repair the foreground image and de-noising.Secondly, the head-shoulder detection for single feature recognizable head-shoulder, low andscale diversity issues, this paper proposes the use of histograms of oriented gradients (HOG) andLocal Binary Pattern (LBP) serial integration features to rival shoulder image feature description,fused into the support vector machine, and its training through Bootstrapping way to get the finalhead-shoulder discriminant model, which has improved the detection results.Then, head-shoulder tracking, tracking algorithm used in this paper based on particle filteruses a weighted color histogram as a way to describe the head-shoulder. Use head-shoulder areahistograms and center initialization filter, the realization of human head-shoulder tracking.Finally, for the DSP platform, with the OpenCV library offline training of SVM classifierswere transplanted and optimization of working head-shoulder detection algorithm, to achievesubstantive results.