| Summary: | 碩士 === 國立中正大學 === 電機工程研究所 === 106 === In recent years, due to the development of medical technology, people are more and more long-lived, and proportion of the elderly population has increased year by year. Therefore, the medical care of the elderly is becoming more and more important. We used computer vision and image processing technology to surveil the medical care of the elderly family.
In this paper, we use two cameras to capture different views in the same scene for abnormal behavior events (such as falling down) detection and identification. This paper combines long-term 2D human contour to establish and identify abnormal behavior models with deep learning architecture.
The anomalous behavior recognition algorithm proposed in this paper consists of three main steps. First, we detect 2D human contour from surveillance cameras. Second, we use motion history image (MHI) method and combine long-term sequence actions into several MHI images. Finally, we use deep learning technology (CNN and CNN + LSTM architectures) to recognize the abnormal behavior of MHI sequences. This method not only recognizes the motions of walking, standing, falling down, but recognizes rising to avoid excessive false alarms.
This paper compared different segments of the abnormal behavior identification. Due to the falling frames are about 3-5 frames, the abnormal behavior (falling down) recognition rate will be low if we didn't divide the sequences in long-time recognition. We compared the performance of abnormal behaviors that used different training methods and different ground truth (GT) of the same deep learning architecture (CNN + LSTM architectures). We also compared the performance that used different architectures (CNN architectures and CNN + LSTM architectures) and single/double cameras. In this paper, we experienced the method that used two-stage training and GT was the majority decision. We got the accuracy of abnormal behaviors identification is 97.66%.
|
|---|
