Vehicle And Pedestrian Detection In Surveillance Video

Vehicle and pedestrian detection in surveillance video is a fundamental topic in IV (Intelligent Video) process. Nowadays, it has become the R&D focus for many research institutes and companies with a broad market prospect. However, it is still a great challenge to achieve robust and accurate vehicle and pedestrian detection, especially in various dynamic scenes and different crowded situations.Based on the state-of-the-art of object detection, this thesis proposes a framework of vehicle and pedestrian detection. The major innovations in this work are enumerated as follows:1. The existing foreground methods are easily affected by dynamic background, generally resulting in inaccurate foreground extraction. For the purpose of solving this problem, a novel algorithm for foreground detection is proposed by combining tempo-spatial consistency into conventional Gaussian Mixture Model and segmenting foreground with region growth method. Further, color truncated cone is combined with LBP (Local Binary Pattern) texture description to remove the shadows and well save the dark color objects.2. Most of the vehicle and pedestrian detection methods are specially designed to a given camera angle, fixed road regions and indicated illumination conditions. To overcome the detection problem in extensive surveillance configurations, novel features are proposed for vehicle and pedestrian detection in road surveillance video sequences. These features are trained in SVM (Support Vector Machine) model, and compared with wavelet coefficient features and block LBP features in both computation complexity and detection accuracy.3. In order to segment each pedestrian from a crowded scene, a novel cascade pedestrian detection approach is proposed. It makes the initial hypothesis of pedestrian region by analyzing contour and foreground information. Then, HoG-based detection method is applied to make the final verification. This approach can improve the speed and precision of pedestrian detection simultaneously, which provides sufficient information for following event detection task.The above-mentioned works are tested thoroughly on extensive real surveillance video contents. It demonstrated the proposed algorithm can detect vehicle and pedestrian efficiently in complex situations. Moreover, the event detection system developed on the proposed framework has helped our group gain two best results in TrecVid 2008.