Research On Key Technologies Of Multi-target Tracking In Complex Scenes Based On Video

Multi-object tracking based on video is a key problem in the field of computer vision, which is very important in the field of intelligent transportation, smart city, virtual reality, sports event analysis and so on. Object detection and data association are two important steps of multi-object tracking in complex scenes. Since the target may be occlused by others, there will be false detection results and missing detection results. If detection results are in big error, the results obtained by the data association will be also in big error. Therefore, this paper focuses on the object detection and data association of multi-object tracking method, and the specific work is as follows:In order to get the accurate target location, this paper studies a reliable method for multi-object detection. A fast target region prediction algorithm based on halftone feature and structured output support vector machine is proposed, which is used to predict each target in the current frame. Execute upper layer detection on this basis, which is extracting the target interesting area from the target area of prediction and foreground area extracted by VIBE detection algorithm. On the one hand, VIBE algorithm cannot detect a stationary target in a period of time, upper layer detection can make up the missing detection. On the other hand, it also significantly reduces detection area for the lower layer detection, and accelerates the detection speed. And then, use lower layer detection in the extraction of the interesting region. Firstly, DPM detection algorithm is used to get the target possible location, and fuse the target area with the prediction,then the location information of all targets is obtained. Experiments show that the hierarchical detection method proposed in this paper can effectively overcome the missed and false detection, and improve the detection accuracy.In order to ensure the reliability of coordinate association, this paper adopts multi-object tracking method based on hierarchical data association. Fristly, we propose a lower association method to carry on the correlation to the target coordinate in the short time, and generate a relatively short trajectory. And then, a object appearence feature based on multi-feature fusion is studied, and the feature pool for each object is established; and on this basis, associate short trajectory by conditional random field model based on the target appearance and motion information. Finally,a feature pool updating method is proposed to ensure the efficiency of the online updating appearance model and data association. The experimental results show that the method can achieve better multi-object tracking performance of accuracy and speed in complex scenes.