Research On Moving Object Detecting,Locating And Tracking

Object tracking is one of the most important branches in the computer vision field. It is used to analyze and manage video sequences to obtain the accurate orientations of moving objects based on their characteristics. It provides useful information and technical support for research and applications, and is widely used in automatic surveillance, human-machine interface, medical treatment, and etc. This paper focuses on the real-time multi-object tracking whose the background is relatively still.The main work in this paper is as following:1. Object Detection: we get the original image sequences which include moving objects by video. In this paper, we review three classic detection techniques: frame difference,optical flow technique and background subtraction, and analyze the advantages and disadvantages of them. We discuss a detection algorithm based on pixel level detecting and frame level detecting in details. In the algorithm, we build the background modeling based on the considering of the pixels whose values do not change sharply for a period of time as the background pixels. The foreground image which is expressed as a binary image is acquired by the difference between the current image and the background model. We analyze the characters of this algorithm by experiments and discuss how to determine the values of parameters.2. Noise Elimination and Object Locating: the opening operation and closing operation are used to remove the noise and fill up the "hole", or "fossa" in the object to decrease the computation. Moreover, we review two classic locating techniques: projection analysis and connected area analysis. Projection analyzing method is effective in the situation that there is only one object in the scene. The connected area analyzing regards several nearby objects as one blob. This paper presents an improved projection analysis method. It could be used in the situation that the scenes include many persons. Experiments show that the proposed algorithm is effective.3. Object Tracking: Combining the improved projection analysismethod, Kalman filter is used to track the multi-human in thispaper. A linear tracking model is designed and the values of statefunction, constant matrix are determined at first based on theKalman theory. Then, the original value of the state vector andpredicted error covariance matrix is determined logically. Thetracking task is accomplished by recursion and matching ofKalman filter. When the humans are occluded deeply or lost in thecase that the improved projection analysis can not segment andlocate each person, the prediction mechanism of Kalman is usedto resolve it. The experimental results show that the algorithm canbe used to detect,locate and track moving humans correctly.The main innovations of this paper are: we analyze the projection ofmulti-human binary foreground image, and find that the horizontalcoordinates of calvarias are equal to the abscissas of local wave crests invertical projection; and the statures of the person (the vertical distance ofcalvarias to feet) are equal to the ordinate of local wave crests. Based onit, a tracking algorithm combined of the improved projection analysisalgorithm and Kalman filter is presented. Satisfied results are achieved inthe real-time multi-human tracking experiments, and we have proved thepresented algorithm has the characters as follows: 1. In the situation, that there are multi-human in the scene, multi-human entering into the scene at the same time and multi-humans are in low occlusion, the improved projection analysis method can segment and locate each person effectively.2. The algorithm analyzes the vertical projection of the binary moving foreground image to segment and locate the targets. The vertical projection is the accumulation of black pixels on each vertical line, so neither the small noise, the "hole" on target, nor the splitting of the foreground target will affect the results which means the algorithm is robust.3. Locations of calvarias correspond to the local wave crests of projection. Even in the condition that the heads are not occluded, or the bodies are merging together, the algorithm can segment and locate them correctly.4. The algorithm track the objects based on the Kalman prediction mechanism. The target lost because of 'bad' background subtraction or deep occlusion can still be tracked correctly. So, it guarantees the continuum of trajectory and stabilization of tracking.