Research And Application On Image Moving Target Detection And Tracking Algorithm

With the construction of smart city and the serious public security situation, Intelligent video surveillance technology plays an increasingly important role in intelligent transportation, intelligence community, public safety and other aspects. Target detection and tracking technology which is the basis for intelligent video surveillance, will directly affect the results of subsequent data processing of intelligent video surveillance. The existing target detection and tracking algorithms have been able to meet the basic requirements of image processing for accuracy and real-time under normal conditions. However, if the appearance of target is very similar, these algorithms can not guarantee the accuracy and real-time processing at the same time, so we study the target detection and tracking algorithms under the above-mentioned complex conditions to improve the speed and accuracy of target detection and tracking in this paper.First, we optimize the two main stages in the cascade Deformable Part Model, namely model matching and cascade part pruning. In the model matching stage, we split the HOG feature variation into foreground and background using SLIC superpixels to improve the detection accuracy. To prun unpromising hypotheses more efficiently, we use the neighboring regions information to improve the pruning criteria in the cascade DPM. When compared to other methods, the experimental results demonstrate that the proposed detection method improved the average detection accuracy on 20 target categories by 1.6%. Detection accuracy on some categories increase by 6% with no obvious decline on detection speed.Second, due to most Network flow tracking algorithm is based on the appearance of the target, it is prone to target ID switching when the target appearance is very similar. We optimize the procedure to generate track segments and compute similarity. We use metric learning to gain reliable tracklet and use tracklet dynamics to handle the situation where the moving targets have very similar appearance. The experimental results show that our tracking method can handle the complicated tracking situation and can improve the average tracking accuracy by 10% with no obvious decline on tracking speed.Third, we implement a prototype system which integrated the proposed detection and tracking method. Our system is developed in the Windows operating system and the VC++ compiler environment. We also called the Intel’s Open CV for the video pre-processing, analysis and moving target detection and tracking.