Research On Human Target Detection Algorithm For Home Nursing

With the development of economy, the trend of aging in China has been increasing, it causes a lot of family tragedy for the elderly people living alone. Research on behavior understanding in the home environment is expected to solve the problem of elderly care services for elderly people living alone, and the detection of human target is the basis of behavior understanding. Therefore, it has important significance to study the human target detection algorithm in the family environment. In this paper, aiming at the research difficulties in the family environment, such as occlusion, illumination change, scale, different posture, and the rapid detection, we research the detection algorithm and the specific detection method, for improving the reliability of the human target detection algorithm in the intelligent nursing care of elderly people living alone. The main research work is summarized as follows:For the problem of complex background interference and detection in the home environment, we study the detection algorithm based on feature fusion and cascade Adaboost. For the problem of leak detection and the fault detection caused by complex background of family, we analysis the anti-interference method of a variety of literature, the features of HOG have strong description ability of contour for the human body, are not sensitive to the change of illumination, and have high detection rate, LBP features have gray scale invariance, faster calculation speed, and can depict the texture characteristics of flat surface more better, so we use a method of cascade of HOG-LBP features to response to environmental interference. For the problem of rapid detection, we use the integral graph technology to accelerate the HOG features, and do the process of PCA dimensional reduction. In order to improve the running speed, we extract 59 dimensional uniform LBP features. At the same time, we design the cascade Adaboost can fast classification, and the linear SVM is used as weak classifiers to reduce the dimension of HOG features more easily.For the problem of leak detection caused by image scale and human body target of lying down in specific detection, we study the detection method based on multi scale and image rotation. For the problem of leak detection caused by the large human body object relative to the image, we study the method of multiscale image detection. For the problem of multi window output caused by multi scale detection, we study the non-maximum suppression method based on greedy strategy. For the problem of leak detection by human body target of lying down, we propose a method to detect the image rotated 90 degrees clockwise when there is no scanning window to return, and rotating the image back to the original image position after the detection is success.In order to verify the correctness of the above research methods, in this paper, the VS2010 platform is used to train a 15 level Adaboost cascade classifier based on the classic INRIA pedestrian database and the open source computer vision library Opencv, then we evaluated the detection rate, false alarm rate, and detection speed of the classifier with the test sample set. Finally, we test the whole algorithm, and the specific methods of detection and location are also verified by experiments.