Abandoned And Carried Object Detection Algorithm Based On The Multilayer Self-adapting Codebook Model

In recent years, the monitoring camera is widely used in many public places, therefore the staffs need to keep watching multiple monitoring pictures at the same time. This is a great loss on human resource, also the accuracy can’t guarantee. So the intelligent video surveillance system is necessary. However the current existing methods have relatively poor robustness or high computational complexity. This paper proposes a multilayer self-adapting codebook model and also uses the method to detect the carried objects, the abandoned objects and the owner. Concrete realization process description is as follows:Firstly, the key to the video processing is foreground region extraction. In this paper, the foreground extraction algorithm uses the multilayer self-adapting codebook model, which establishes a four layer codebook background model and each layer uses different size of pixel block instead of a single pixel in original algorithm. The computing speed and accuracy are significance improved. The original algorithm which is sensitive to the illumination change, shadow, and dynamic background also has good improvement. The foreground region of the scene can be extracted more accurately.Secondly, the abandoned objects can be detected based on the foreground region extraction result. But according to the characteristics of abandoned objects, we should extract the fixed foreground regions only. This paper establish a multilayer feature codebook model to describe the characteristic of the pixel block, and dynamic adjust the threshold the codeword used based on the variations. Specific to the scene that abandoned objects were blocked repeatedly by the pedestrians, the fixed foreground can be judged by cumulating the times of occurrence. After the abandoned objects obtained, we use the algorithm superpixel tracking to find the location of the abandoned objects in the former scenes. Then the tracking result is used to traverse all the foreground regions of the scene to find the owner efficiently.Finally, carried objects can be detected based on the foreground region extraction result. The human template can be obtained by extract and align the foreground segmentations of the walking pedestrian. Then the human template is matched against a standard human template to find the potential regions. Adopting the markov random field based on the MAP combined with the prior template and the probability distribution of the location of the carried objects, the real position can be segmented correctly.