| Summary: | 碩士 === 國立臺灣科技大學 === 材料科學與工程系 === 100 === This thesis applies image processing techniques to detect a person who falls down and is unable to get up by oneself after fall event. In this thesis, when the event is detected by the detection system, a warning message can be provided to surveillance officers in a computer-based vision system. In the detection system, at first, the motion detection is used to find the moving object (human). Next, we propose three stage judgment rules to detect fall event. In the first stage, orientation variation of the object is used to determine the walking and falling gestures. In the second stage, back propagation neural network (BPNN) with three feature inputs is used to detect fall and stoop or squat situations. In the third stage, we compute orientation variation of the object for detecting disability to get up by oneself and get-up action after lying on the ground. In the experiment, fifty eight video sequences include sixteen normal events (walk and stoop or squat) and forty two fall events (forward fall and backward fall) for the first stage and second stage judgment rules, and sixteen video sequences include eight for disability to get up by oneself after fall event and eight for get-up action from the ground. The recognition rates of two experiments are 94.8% and 100%, respectively. The results show that the proposed fall detection system can do good job in recognizing fall events.
|
|---|
