Object Detection And Analysis In Surveillance Scenes

As a hot spot in the field of video surveillance, pedestrian detection and behavior analysis are becoming more and more popular. One of the prerequisites of behavior analysis is moving object detection, followed with pedestrian counting. In the area of intelligent transportation and public security, it is necessary to detect and analyze the objects in the surveillance scenes with computer vision and pattern recognition techniques. Monitoring manually is huge work and the results would be inaccurate due to carelessness. However, the state of the art algorithms of crowd counting are mainly based on tracking or regression. Visual tracking is computationally expensive and tracking multiple targets accurately and robustly is still a challenging problem. The regression based algorithms are able to deal with crowded scenes well, but hard to achieve real-time.We not only introduced object detection methods based on background modelling, such as Gaussian mixture model and first-frame difference method, but also proposed an image based object detection method. Our method is more applicable in surveillance scenes compared with traditional detection methods. We detect targets in a region of interest and control the size of sliding window. By reducing the sliding area and the types of sliding windows, our detecting algorithm is more efficient and can help the surveillance system to reach real time.For pedestrian counting based on tracking, we proposed an algorithm to analyze each tracking target based on their gait features. The proposed algorithm is able to distinguish pedestrians and non-pedestrian objects, such as suitcases and bags. In the object analysis algorithm based on gait features, background modelling is applied to detect the foreground moving objects. Gait features are extracted from each foreground object in order to judge whether the foreground is a pedestrian or not. This classification algorithm is based on the fact that the gait features of pedestrians are usually dynamic while the features of non-pedestrian objects are mostly static. As a result, the counting accuracy of tracking based methods is improved.In this paper, we also proposed a crowd counting method based on spatial and temporal analysis. Combining the tracking based and regression based methods, our algorithm can detect and count pedestrians in crowded scenes and operate in real time. First a temporal slice image is obtained which contains spatial and temporal information. Then the foreground parts in the temporal slice image are classified by support vector machine and later clustered. After analyzing the spatial and temporal information cyclically, the number of pedestrians as well as the moving directions of each pedestrian is obtained. The surveillance system can achieve real time by analyzing the human behavior cyclically.Experimental results show that the object analysis algorithm based on gait features can accurately distinguish pedestrian and non-pedestrian objects, and count the number of pedestrians in non-crowded scenes precisely. For the method of spatial and temporal analysis, extensive experiments reveal that both the number and the moving directions of pedestrians can be estimated accurately and efficiently.