Pedestrian detection and identification is one of the most important research fields in computer vision and pattern recognition. It also has broad application prospects and market value in visual surveillance, virtual reality, human-computer interaction and driver assistance systems. Pedestrian detection and identification is the key point for pedestrian tracking, behavior analysis and alarm process-ing. Although the existing pedestrian detection and identification algorithms has got good performances on various pedestrian databases, but most of them still can not meet the requirements in real-time monitoring system due to the high computational complexity. To solve this problem, this thesis focuses on real-time pedestrian detection and identification for visual surveillance. The main work and contributions are as follows: 1. This thesis first introduces the existing pedestrian detection and identifica-tion methods,such as the segmentation region of interest, rigid template based and part based model for pedestrian detection. In the identification phase, we mainly introduce color features, texture features and shape fea-tures based pedestrian identification methods. 2. Because of the existing visual surveillance systems mainly based on fixed cameras, according to this feature and requirements for real-time, we pro-pose a boosted cascade based pedestrian detection method for fixed camera scenes. We use the segmentation of region of interest and scene learning model to reduce the subwindow search space. The Multi-scale HOG feature and boosted cascade classifier are applied to determine whether the window contains a pedestrian or not. This method obviously improves the detection speed of the boosted cascade method without decreasing the accuracy. 3. Over the last few years many object detection methods were proposed. With better performance, the efficiency becomes more and more important. But existing methods are based on high dimension features, and the computa-tion complexity of high dimension features directly affects the detection effi-ciency. We propose a partial least squares based subwindow search method for pedestrian detection which obviously reduces the number of search win-dows from the search strategy. Firstly, a sparse search is implemented to find all the possible locations containing parts of a pedestrian. Then a pre-learned Partial Least Squares regression model is applied to estimate the displacements of the subwindows to guide th...
修改评论