Research On Target Detection Recognition And Tracking Based On Surveillance Video

Video monitoring has become an important part of the modern safety system, no matter road transportation, warehouse, workshop, fence in school, bank security,security company have adopted video monitoring as an important tool for guard against security. However, the intelligent computer vision technology which embedded in the video monitoring is becoming more and more outstanding. For example, target recognition, cross-border identification, trajectory-tracking, lost-property recognition,license plate recognition, forecast alarm and other intelligent understanding technology become smarter. Meanwhile, real-time tracking of specific targets is also an important content of video monitoring.The main topic of this paper is video monitoring of intelligent understanding technology, including video of multiple target detection, recognition and tracking, etc.We propose the super-segmentation method which combining multi-level-color space with random selection, it can satisfy the coverage of the multiple target segmentation and intelligent tagging target in video scene. The Visual dictionary classification,which is based on SURF feature, can significantly improve the recognition accuracy. Proposed the multi-target tracking method, which is based on similar feature and relative movement, can achieve real-time tracking multiple targets.The main research contents and contributions in this paper:Firstly, the super segmentation method based on random selection and multi-level color space is proposed. In the Complex multi-target scenarios, the traditional methods require to scan the entire video and pinpoint the targets. This method randomly selected1,000 to 10,000 locations in a multi-level color space, located the image area that can described the target category. The method does not need to scan the entire video sequence, and not only significantly reduce the time overhead of object detection, but also ensure full coverage of target segmentation(recall), provide sufficient information for effective follow-target recognition.Secondly, the visual dictionary classification method which based on SURF features is proposed. For the low recognition rate problem of bag-of-words(referred to BOW), we create a visual vocabulary using SURF features, which significantly improvethe classification results. Using the SURF-visual dictionary classification method for ordinary scene classification, can achieve intelligent targets understanding and accurate classification scene. Combining the SURF-visual dictionary and feature space visualization can achieve a certain degree classified of dark scenes.The last, multi-target tracking method is proposed based on a combination of similar characteristics and relative displacement. For real-time video surveillance, using the low-rank matrix approximation to solve the problem of unavailable appearance for similar goals, combined with similar characteristics and generalized linear distribution,can achieve real-time tracking of multiple targets. Using objective information recovery "indicator" vector, can effectively solve the target loss and recovery.