Research And Implementation Of Feature Fusion Tracking Algorithm Based On Nearest Neighbor Decision

With the development of people's life quality, the intelligent video surveillance system is needed in the fields of national security, public security, industrial production and service of daily life. Even though the intelligent video surveillance system may reduce the workload of staffs and solve problems in manual monitoring, it can not meet the needs of people at present. That is to say, on the premise of hardware resources and labor cost, improving the performance of video monitoring system and trying to adapt all the monitoring environment is very significant.Traditional target tracking algorithms based on single feature track well under simple environment, however in practice, the environment is complex and there are multiple people appearing in it. Then, tracking object is easy to be influenced with other moving objects or background. Aiming at the shortcomings of the single feature, the thesis is based on multiple features fusion theory, and is applied to the human target tracking.The main tasks are concluded as follows.(1) Moving target detection is an important foundation of target tracking, Integrity and accuracy of target area are related to accuracy of the subsequent tracking. Extracting the complete moving target area is extracted by fixed video surveillance of single camera. In order to carry out the complete moving target area, improved three-frame differential method by void filling is used in this thesis, which performs better in effectiveness and accuracy of obtained complete moving target area than the classical GMM.(2) In view of tracking the selected object in complex environment with multiple people, accuracy and efficiency of the algorithm are considering mainly. Multiple features fusion can improve accuracy and adaptability to different environments of the tracking algorithm, but computational complexity of the algorithm also increases. In order to solve the problem, due to the high requirement for time, the nearest neighbor rule is introduced in this thesis to find the nearest candidate target in the following frame. In order to solve the problems such as color interference during target tracking, the method in this thesis is to extract the color and geometry feature from targets and candidate targets. Harris corner matching has been applied to solve problems like target occlusion and so on. According to experimental verification, the proposed algorithm has improved the tracking accuracy and robustness.(3) Real-time target tracking system is built on experiment by using the Hikvision SDK, which can obtain the real-time video streaming. The system built in this thesis can track specified target in various and complicated environment. Comparing with the traditional tracking methods, through experiments, the method in this thesis improves the accuracy and efficiency of tracking target obviously.