Indoor Scene Human Body Tracking And Activity Analysis

The work of this thesis is a part of "multi-pattern intelligent visual perception and mobile 3G network based empty-nest elderly monitoring system" project, which aims to recognize indoor empty-nest elderly activities, and alarm the dangerous accidents. By improving and integrating the algorithms of foreground detection, human body tracking and activity analysis, a vision based automatic empty-nest elderly monitoring solution is implemented, which has a wide range of applications. The monitoring solution has a PC client which can do robust tracking and activity analysis on empty-nest elderly. Then the PC client uploads the activity analysis results to the remote server, which finally pushes the message to the Android App client.The main works are summarized as follows:1. Moving object detection and multi-human body tracking. In indoor scene moving object detection research, three frequently used algorithms were compared: DOF(Dense Optical Flow), ViBE(Visual Background Extractor) and GMM(Gaussian Mixture Model). Finally the GMM based algorithm was chosen, because it is well-balanced in all aspects. Moreover, the GMM based algorithm was improved with shadow detection to remove shadow, and morphological method to cluster human body foreground Blobs. In the real environment, the classical HOG(Histogram of Oriented Gradient) feature based pedestrian detection method has high error rate and high computation cost, which is difficult to be adopted in real-time surveillance system. So the HOG based pedestrian detection is carried out only on tracked foreground Blobs area, which had greatly improved the accuracy and speed of person detection. Then a simple Blob Tracking algorithm was implemented, which is mainly based on the Blob's geometric features between adjacent frames. The foreground detection method can not work when an elder is standing or sitting. Moreover the foreground detection based Blob Tracking can not work when the multi-person occlusion happens. To overcome these shortcomings, the Blob Tracking method was improved with KLT(Kanade-Lucas-Tomasi)Tracker, which resulted in a robust human body tracking.2. Action feature extraction and activity classification. Compared to binocular camera, the monocular camera can not get the real velocity and position of the tracked person due to lack of scene depth, which makes it difficult to analyze activities. By supposing the indoor ground as a plane, the ground geometric information was calibrated, and PPT(Perspective Projection Transformation) was applied to get the real velocity and position of a tracked person, which had greatly improved the accuracy of activity analysis. Because there are no public indoor video datasets with geometric calibration, a video dataset was recorded which contains five activities: walking, running, sitting, standing, and lying. The activity features consist of velocity, position and geometric information of a tracked human body in continous frames. Then GBM(Gradient Boosting Machine) was trained with 1185 activity samples extracted from the dataset, and the recognition result reached 96.21%. Finally I had realized the alarm of dangerous accidents of empty-nest elderly.