The Research And Implementation Of Fast Object Detection In Crowded Video Scenes

Nowadays, the number of video surveillance cameras for public security in a modern city has generally reached more than ten thousands. Aimed at pedestrians and vehicles, object detection in city video surveillance systems is one of the hottest research areas in computer vision and pattern recognition. However, accurate and fast object detection in the crowds, especially for pedestrians, is one of the most difficult challenges in this area.In this thesis, a survey on the existing object detection algorithms is conducted. Deformable Part-based Model(DPM) gets a better performance for occlusions in objects in the crowds. But DPM uses global sliding windows and feature pyramids, the high algorithm complexity, which leads to low real time performance, makes it difficult to be applied to the actual intelligent video analysis system.Designed for the actual intelligent video analysis system, a hierarchy detection method using the “coarse to fine” strategy is proposed. Firstly, symmetric difference algorithm is applied to quickly locate the potential object area in the given video clip. Then, a precise detection is conducted in the potential area using the DPM detector. In this way, the algorithm complexity is reduced to satisfy the need of both precision and time in certain crowd situations. Meanwhile, a fast coarse object locating method using BING+DPM is also proposed to realize the requirement of real time object locating in crowd scenes.The proposed algorithms are evaluated on pedestrian datasets like INRIA, PETS2006, etc. Precision ratio and recall ratio is obtained to plot the PR curve. The existing videos in the practical system platform of the engineering center is also used for the evaluation. The result shows that, due to the performance of BING algorithm, the precision of BING + DPM algorithm is lower than the symmetric difference + cascade DPM algorithm, but the high detection speed makes BING + DPM algorithm useful in the real time application of the video analysis system. The symmetric difference + cascade DPM algorithm has a better precision ratio, which is useful for the actual requirement of the video analysis system in certain situations. The dynamic combination of the two algorithms above fulfills both the accuracy and real-time requirements of object detection in certain crowd videos.