Research On Object Detection And Tracking

At present, Intelligent Video Surveillance is becoming a hot research field in Computer Vision. In the area of Intelligent Video Surveillance, generally interested in pedestrians, Detecting and Tracking of Pedestrians has naturally become the most important research emphasis. This paper makes some improvements in the algorithms of Pedestrian Detection mainly based on HOG and puts forward its own proposals with regarding to the deficiencies of HOG features. Accordingly, within the area of Pedestrian Detection, the paper advances an efficient algorithm of Multi-object Tracking on the basis of discussing inadequacies of popular tracking algorithms.In Object Detection, this paper proposes an improved Pedestrian Detection algorithm based on HOG features. First, we introduce Integral Histogram of Oriented Gradient, which effectively shortens the computing time of HOG feature. Then, we analyze eigenvectors composed by every Cell with respect to HOG features with the method of PCA. In addition, a Multi-scale HOG feature is brought about and algorithm of Adaboost is used to select features, for the shortages that HOG block is simplex in size and fails to express some features in part of pedestrians on a semantic level. Experiments show that Multi-scale HOG feature, compared with HOG, makes great improvement both in detection performance and detection time.In Object tracking, with idea of Network Flow popular in graph theory introduced, this paper proposes the algorithm of Multi-object Tracking, which is able to utilize EOM model to add hypothetical occluding targets when objects meet with occlusions. The problem of global target association is conversed into that of Min Cost Max Flow. Thus, those existing algorithms of Network Flow could be used to solve such problems. Through enough experiments, we find that tracking problems with occlusions could be resolved efficiently using the algorithm.Finally, this paper designs a Detecting and Tracking Surveillance System in order to verify the feasibility and effectiveness of algorithms discussed above.