Research On Human Fall Detection Based On Deep Learning

Fall is the leading cause of accidental death of the elderly,and fall behavior detection has important application value in maintaining the safety of the elderly.In this paper,the research of human fall detection under surveillance video is the principal direction.Based on computer vision,the current popular deep learning technology is applied to the modeling and detecting the human fall behavior.The main tasks are as follows:1)There are three existing methods of human fall detection were researched,including wearable device method,environmental perception method and computer vision method.And advantages,disadvantages and applicable scenarios of different methods were compared and analyzed.2)A human-falling-down detection model based on the keypoints of human contour and the LSTM neural network is proposed.This model proposes a new method for extracting contour keypoints based on the center of mass,it extracts thirty contour keypoint coordinates and combines centroid coordinates as human features.With the model continuously detects multiple frames of images to obtain human feature sequences,it divides the feature sequences into X and Y coordinate sequences,then it imputs the coordinates respectively into two LSTM neural networks to extract sequence characteristics and obtains the final results by inputing two LSTM hidden layer output vectors into a fully connected layer.This research applies public dataset MuHAVi-MAS14 for experiment and its recognition rate of human-falling-down detection can reach up to 99%,and it also adds experiment of multi-actions recognition based on the dataset and the recognition rate can over 90%,proving the validity of the model.3)A human-falling-down detection model based on the human skeleton keypoints and the LSTM neural network is proposed.The model detects the skeleton keypoints of the continuous multi-frame human body by Alphapose as human feature,then it use LSTM neural networks to extracted sequential features from human feature sequence,the final classification results are obtained by a fully connected layer.This research uses public dataset MuHAVi-MAS17,Le2i and MCFD to execute this experiment,and the fall detection recognition rates reached 92%,99% and 92% respectively.The results show that this model has pretty good universality for multiple views,multiple scenes and multiple poses of falling.4)A human-falling-down detection model based on YOLOv3 and lightweight convolution network is proposed.This model detects the human body of the continuous multi-frame images by YOLOv3,then it use the MobileNetv3 network to extract the human features,and it fuses continuous feature vectors of multiple frames next.Finally,the classification is achieved through the fully connected layer.The evaluation experiments were performed on the public dataset Le2i and MCFD,and the fall detection recognition rates reached 98% and 95% respectively.The results show that the model has a high recognition rate and good comprehensive performance.