Fall Action Recognition Based On Deep Learning

Falling action recognition has broad application prospects and great economic value in the field of medical monitoring field.In recent years,with the continuous development of computer image processing and artificial intelligence technology,the fall action recognition technology has also made significant progress.Visual-based fall action identifies multiple problems such as environmental disturbances,actional complexity,and perspective changes.This research studies the these issues of fall action recognition.In view of the problem that traditional methods need to extract features manually,the fall action based on a deep learning was used in this research.Convolutional neural networks can learn data features autonomously.This method could reduced the complexity of the old algorithm and improved the accuracy of fall action recognition.In this research,the data enhancement method was used to augment the initial dataset.It could avoid over-fitting of the model.In view of the insufficient information fusion and small target missed detection between the network layers of the original SSD target detection algorithm,the feature information of different network layers was merged to construct the feature pyramid.By increasing the context information of the network model to improve detection accuracy.In order to meet the real-time requirements of fall detection,this research cuts out the redundant parameters existing in the network,reduces the amount of calculation to improve the detection speed.Because the original SSD target detection algorithm was susceptible to fall similarity behavior and background factors.Aiming at this problem,a fall action recognition method based on pose estimation is adopted in this research.This method extracts the joint point information from the video sequence and constructs the skeleton spatial-temporal graph model.By adjusting the graph convolution network model,the optimal model was obtained.The experimental results show that the algorithm has good performance in detection accuracy and detection speed,and has good generalization in human action dataset.Finally,we summarized the research work and looked forward to its development.