Moving Target Detection And Tracking Algorithm Based On Compressed Domain Is Studied

It is always difficult to detect and track moving objects in videosand thus making it a hot spot. The purpose is to detect and segment themoving objects which people are interested in. This technique is widelyused in video indexing, intelligent monitoring and pattern recognition.The majority researches of moving object detection and trackingare applied in pixel domain. However, there is great amount ofinformation in pixel domain, unable to meet the requirements forreal-time usage. On the other hand, with the rapid development ofmultimedia technology and that of video compression standard, videois generally stored or transmitted in the compressed domain. Soresearch on moving object detection and tracking in compresseddomain is of great importance.In compressed domain it only requires partial decompression formoving object detection and tracking. Compared to completelydecompression, it takes less time. In addition, in compressed domain weusually regard a block as a unit, while the number of pixels in videos ismuch more than that, which makes it much faster to processinformation in compressed area.This paper mainly researches on moving object detection,segmentation and tracking algorithm in H.264compressed domain. Firstof all, a method of threshold-based object detection is proposed incompressed domain. After studying the codec information of H.264, weget the motion vectors, which are extracted when it is onlyincompletely decoded. Then pretreatment is applied for motion vectorsand moving objects are detected according to temporal correlation,which is based on threshold and static video sequences.Secondly, this paper presents a moving object segmentationmethod based on particle swarm clustering algorithm, applying improved K-PSO clustering algorithm to classify the motion vectors afterpretreatment. Then information in macro block layer is introduced todifferentiate the noise and moving part.Finally, a method on moving object tracking is proposed incompressed domain. It regards direction angle of motion vectors assystem measurement, applying particle filter to the algorithm toestimate the true state of current object. This algorithm is less affectedby noise, but is sensitive to the moving direction of detected object.