Research On Moving Object Detection And Tracking Algorithm Based On Adaptive Hybrid Model

With the development of computer vision and image processing, intelligent video surveillance technology has also made considerably progress. As the core of intelligent video surveillance systems, making the computer analyzes the video automatically and accurately, has increasingly become the focus of many scholars in the related fields.The detecting results of moving objects make important influence to target tracking, classifying, and behavior recognition. Based on the analysis for existing LBSP(Local Binary Similarity Pattern) algorithm which uses fixed threshold for all pixels to calculate LBSP values resulting that the adaptation of background model is poor, this paper uses the standard deviation of difference values between the center pixel and its neighboring pixels as an adaptive threshold value for each pixel to calculate its LBSP. Then build a mixed background which blends pixel's texture and brightness information and start detecting moving objects. Last, classify the current pixel based on the result of comparing the current pixel and its corresponding mixture background model, and update the background model randomly based on the classification result. Experimental results show that this method not only achieved better test results under normal ambient environment, and can effectively reduce the disturbance maked by dynamic background, lighting changing and so on, and can improve the accuracy of test results.After get the moving objects, establish a target template by extracting the effective feature of interesting target. Then determine the location of the interesting target(if any) in subsequent frames. Extract Shape Context information of the moving target which is got by the above-mentioned adaptive mixture model to build a target template. Then in subsequent frames, to extract all Shape Context information of the detected objects to establish the candidate templates, then match the candidate templates with the target templates, and filter out the candidate templates with the highest similarity as the target in current frame, then update target template. Experimental results show that the results of the proposed detection method can remain true shape information of the objects, and the proposed target tracking method can get better results for tracking non-rigid and rigid objects.