Research On Moving Object Detection And Action Analysis

Video analysis is an important tool for active vision information mining and plays an important role in object detection, object recognition and abnormal behavior detection. With the increasing quantity of the surveillance data, intelligence video analysis turns out to be much more competitive compared with conventional manual moritoring scheme. How to automatically detect interest objects or behaviors in videos and to make an alert for further processing has been an important research task in intelligence surveillance.Moving object detection and recognition is the core content of video analysis, and an important means for image understanding. However, in the surveillance environment, several factors such as lighting changes, shadow interference, video capturing with inclined angle, bad alignment of the recognition targets have negative influences on the performance of object detection and recognition. It is a key problem to extract effective features and to obtain a roubst representation of the target under these complex environments. This thesis designs a series of statistic models and subspace models to solve the key issues of the tasks in video analyis based on the combination of the statistic and subspace methods. The major research content and experimental results are concluded below.(1) Based on the within class maximum principle, this thesis proposes a frame set division method as a preprocessing step for low-rank decomposition in moving object detection which can greatly improves the average detection accuracy of the Pets 2006 dataset etc. Current low-rank decomposition methods rely too much on the subspace models for background modeling and neglect the statistical characteristics such as the distribution of the locations of the moving objects for background subtraction. This work is able to obtain the frame subsets whose distribution of the sparse noise is beneficial to the seperation of the foreground and background after frame set division. However, the subsets suffer from a small scale problem. To this end, we propose to augment each subset by using the frames with slight motion intensities, which increases the number of the genuine background pixels. This work explores an effective way for the combination of statistic methods and subspace methods in background subtraction. The proposed method not only makes advantages of both types of methods but also achieves stabler robustness under the complex surveillance environment such as lighting changes and severe occlusion.(2) In this work, we propose an unsupervised optimal feature selection method to integrate dimensionality reduction, sparse representation, joint sparse feature extraction and feature selection as well as classification into a unified objective function for action recognition. Traditional subspace methods usually construct the graph and conduct the classification in two seperated optimization steps. Therefore, the projection matrix derived from the first step is not optimal for classification. Our method optimizes the projection matrix and sparse representation matrix simultaneously so as to select the features which are the most benificial to the classification. For theory analysis, we also prove the convergence of the proposed algorithm and provide a detailed complexity analysis. Finally, we compare our method with other state-of-the-art methods on public human action recognition datasets to validate the effectiveness of the proposed method.(3) This work proposes a multiple direction-based Gaussian modeling method for velocity learning. The proposed method is able to detect the fast moving objects in the surveilance system without any calibration step. In the view of the management, fast moving in the bank is recognized as an abnormal event. The velocities of the moving targets in the video are suffered from non-linear variation because of the projective transformation. We first analyze the relation between the velocities of the people in the image and their real velocities. Then we proposed to learn the means and variances of the velocities of the moving objects in different direction bins. In order to determine the major direction bin, we use the basic Fisher model for optimization. However, we cannot find the cusp at the extremum of this model. To this end, we propose an improved Fisher model to optimize the problem and we obtain the cusp at the extremum of the improved model. After the training phase, to solve the lack of data problem for learning at certain pixels, we proposed a weighted random sample consensus algorithm for parameter fitting in the three dimensional space. The newly designed algorithm tries to estimate the parameters of the lack-of-learning pixels by using the parameters of the pixels with sufficient learning. Finally, we compare our method with other methods on real surveillance videos provided by the bank. Experimental results show that our method can obtain higher detection accuracy and lower false alarm.(4) This thesis proposes a people counting method based on the combination of Ada Boost and multiple features. In our system, the camera is mounted overhead. First, we train a classifier by Ada Boost to detect the interested objects in the video. Then we propose to integrate multiple features to confirm the detection and tracking results. The multiple features include the matched response features, motion intensity features and scale features. Aiming at different application environments, we propose two kinds of tracking scheme. The first one is a nearest neighbor-based tracking scheme which is mainly used in the concise environment with clear vision. The second one is an optimal clustering-based tracking scheme which is mainly used in a much more complicate environment. In order to test the proposed counting method and other methods, we establish a counting dataset with different scenarios. Our method is the best among all the compared methods and the average counting accuracy of our method is above 95%.To sum up, in order to solve the key technologies in moving object detection and action recognition, we extract effective features from videos and make a robust representation of the targets based on the combination of statistic methods and subspace methods. The proposed methods and designed systems work very well on public datasets and real data. Some of our algorithms have been applied to the real-time surveillance system.