Monocular Image Sequences Of The Target Tracking Algorithm

The complexity of environment, such as illumination variety, background changing, shadow, occlusions, makes object tracking a challenge problem. In this thesis, we focus on the human tracking in indoor, discuss the algorithm of moving object tracking, and deal implicitly with shadow and occlusions, which is able to correctly remove the shadow and label people during occlusions.In the phase of object detection, we establish an adaptive Gaussian model for background, which can handle the illumination variety and small motion in background. Then, we obtain the foreground regions by background subtraction. In order to deal with the shadow which is critical for accurate object detection and tracking, we implicitly discuss the feature of shadow in the view of color constancy and photometric invariability, and propose a method for shadow detection. The experimental results show that our proposed method can be able to remove shadow and extract the foreground moving objects accurately.In the phase of human tracking, we adopt different algorithm to deal with the occlusion and non-occlusion. When there is no occlusion among people, we make use of the Kalman filter to predict and estimate the attributes of the centroid of the people in the next frame, achieve the human tracking by tracking the centroid of the people. In order to deal with the occlusions, firstly, we propose a new method in the view of the color distribution, which represents a person by a Gaussian mixture model. We estimate the parameter of Gaussian mixture model by EM algorithm, and then, obtain the human model. When there is occlusion among people, divide the pixels into each people. Lastly, we make use of the spatial information to verify which improve the ratio of segmentation. The results show that our proposed framework can deal with the human tracking successfully, and can be able to label people correctly during occlusions.