Fall Recognition In Open Scenes

As a typical abnormal action behavior,falls occur frequently and usually cause some harm to the human body.And it may even cause economic losses.The falling behaviors' recognition in the open scene does not require the cooperation of the recognized person,and is not limited indoors,so has high practicability.Therefore,t study on the falling behaviors' recognition in open scenes has higher theoretical significance and practical value.This thesis,combining computer vision and deep learning technology,I proposed a key frame selection method and a two-stream neural network.The key frame selection method is based on the image LUV local maximum and Mask-RCNN network.The two-stream neural network is given to recognize the falling behavior of the key frame.finally,the influence of weak illumination is further considered.Improve the Retinex algorithm for image enhancement and then perform falling recognition.The specific content is as follows:(1)In order to eliminate video frames with low correlation of falling behaviors,this thesis proposes a video-key-frame-extraction algorithm which combines LUV local maximum and Mask-RCNN.First,the posture change frame is extracted as the candidate key frame through the local maximum value of LUV;Afterwards,the Mask-RCNN is used for human detection.Then according to the aspect ratio of human body characteristics and movement speed,the videos that are less related to the falling action behavior are further eliminated frame.Experiments and results analysis are carried out on the UR fall detection dataset(URFD),Multiple cameras fall dataset,Le2 i fall detection dataset datasets and real scene of fall video data,verifying the effectiveness and accuracy of the method in this paper.(2)For falling behaviors' recognition,this thesis proposes a two-stream neural network model based on the Mobile Net network.The two-stream network includes both spatial and temporal streams.Taking the key frame RGB Image sequence as input,the spatial stream extracts behavioral contour features in the video.The spatial stream loads the movement history graph of key frames and extracts temporal features.Then the two types of extracted features are fused and analyzed.The proposed method is experimented on the URFD,Multiple cameras fall dataset,and Le2 i fall detection dataset datasets.Compared with single-flow model,3D-CNN model and two-flow model combining CNN and optical flow,The experimental results show that the network model proposed in this paper has improved recognition accuracy,precision and recall rate.(3)Considering that none of the existing fall datasets provide video data in low light conditions,the thesis creates a fall data set based on indoor and outdoor scenes for low light conditions.On this dataset,combined with the improved Retinex algorithm for image enhancement.two-stream network performs fall behavior recognition.The experimental results show that comparing the single-stream model,the 3D-CNN model,and the two-stream model combining CNN and optical flow on the self-built dataset,the recognition accuracy of the network model proposed in this thesis is higher.