Research Of Change Detection In Image Sequence

Change detection is a process of digging out pixel set that change significantly in the sequence of images taken in different time periods for the same scenario. As a basis for computer vision, it is widely used in the fields of anomaly detection, driver assistance, target tracking, and medical diagnosis and treatment. In this paper, the image sequence is the main object of study, change object extraction and background model update method ha ve been studied to improve the detection quality of complex scenes, specify studies are as follows:(1)Since non-parametric model update process is easy to mix in pixels of changing target, and the background changes usually update unsynchronized, this paper proposes a new update method which combines stochastic update and Competition and cooperation update, so that it can describe the dynamic background much more accurately. Traditional sample FIFO updating method and background pixels selected only updating method, they both are unable to meet the accuracy requirements of complex scenario background model. The proposed stochastic update method can comply with the law of randomly residence time of samples in the background model. Competition and cooperation mechanism fully use of the spatial information. Simulation results can not only eliminate false detection quickly, reducing voids occur within the target, but also can handle long changes in the scene efficiently.(2)In complex scenarios, the overall distribution assumed by parameter detection cannot contain all the cases in the scene, lead to higher detection error. Parameter detection depends on the selection of parameters, inflexible and susceptible to the interference from dynamic background, proposed a change detection method based on statistical tests. Use RBJ statistics of goodness of fit test to determine the degree of fit between to-be-detected pixels and background model. Separate the target according to the fit, update the background according to split Gaussian principle to reduce the cumulative error in background model. Simulation results show that when confronted dynamic background, ca mera Jitter and other complex scenes, the algorithm can effectively suppress the dynamic interference, have good recall and overall performance indicators.(3)Because of changes in the target and the similar background color will cause target missed or incomplete detection problem, proposed belief propagation and energy optimization method used for target detection. Belief propagation algorithm can achieve information exchange between the pixels among temporal and spatial domain. Then calculate the amount of information for each pixel, so it is possible to extract the changing target by comparing correlation between the to-be-detected pixels and background pixels. According to the slowly varying characteristics of background in the time domain, record the change of background in time matrix and adjust the time information in order to improve the accuracy of the time-domain information. Simulation results show that the belief propagation algorithm can effective transfer information between pixels, the algorithm can adjust to the complex scene, track changes in the background and detect change target in the complex scenes effectively.