Video Object Tracking In Multi-camera Environment

Video tracking system plays an important role today because of the rapid development of science technology. Its development prospect is extremely optimistic and it has been used in many fields. This system is aimed to extract and track the interested video objects in video sequences, to analyse behavior, as well as to alarm for abnormal behavior, and then complete the daily management and protect the security of tracked places. Currently the emphasis in this research field is that how to find an automatic, real-time, effective and accurate tracking method.This paper mainly studies the tracking problem based on multi-camera environment. It includes mainly the extraction and the tracking problems based on single camera environment and the continuous tracking problem based on multi-camera environment. In the aspect of video object extraction, our paper proposes an adaptive object extraction algorithm based on multiple cumulative frame differential method through combining cumulative frame differential method with twice frame differential method. As to the multi-object tracking in complex scene, we propose the method of object tracking using multi-feature fusion. For continuous tracking in multi-camera environment, we propose the continuous tracking algorithm based on multi-feature fusion, which is used in multiple non-overlapping cameras environment.The video object extraction method proposed by us is to detect changed region using cumulative frame differential method between the current video frame respectively with the former several frames and latter several frames first, and then take the intersection of them. At last we get the final video object. Here, we can choose adaptively the number of the frames of cumulative frame differential method through estimating whether the motion information of current cumulative frame differential image is sufficient or not.In the object tracking algorithm proposed by us, we store the object information of former several frames in former object set first. When there are objects both in the current frame and in the previous frame, we use gray model to forecast the position of objects in the current frame, and then fuse position feature, brightness statistics feature and shape feature to complete object matching. Then we judge object state of motion. We judge the type of occlusion through occlusion disposing when the object occlusion exists in the current frame. If the occlusion is mild, we use template matching based on weighted histogram. When there are objects in the current frame and no object in the previous frame, all of the objects in the current frame are considered as the new objects. When there are objects in the previous frame and no object in the current frame, we consider all of the objects of the previous frame disappear in the current frame. Then we judge whether the disappearing object is temporary through calculating the number of continuous disappearing frames. Finally, we label the objects in the current frame, and update former object set.The proposed object continuous tracking algorithm is to adjust the histograms of matched objects in the current camera horizon and in the adjacent camera horizon respectively first. Then the first matching is employed by using luminance statistics feature and shape feature. If the similarity of the first matching is larger than the given threshold, the second matching based on brightness distribution feature is employed. Finally, we judge whether the current object is the new object or the disappeared object of the adjacent camera horizon through whether experiencing the second match and the value of second matching similarity.From the results, we can see the proposed object extraction method can extract video objects accurately and reduce computational complexity; the proposed object tracking method can complete the multi-object tracking in different scenes; the proposed object continuous tracking algorithm is real-time and precise in continuous tracking of objects in multiple non-overlapping cameras environment.