| Falling has become one of the primary health threats for the elderly in the context of increasing global aging.With the continuous development of safe cities,intelligent transportation and other construction in China,the integration of machine vision technology into video surveillance systems has become a hot research issue.Due to the rapid rise of deep learning,the use of convolutional neural networks for image classification has achieved good results.This paper explored the methods of machine learning and deep learning,and detected the fall of the human body in the surveillance scenes.The main work and contributions are as follows:1.A real-time fall detection method based on key multi-feature fusion of human body was proposed.Firstly,the Retinex image enhancement algorithm based on HSV space was used to enhance the images,improved the contrast and brightness of the dark area,and improved the image quality.Secondly,the method of combining Gaussian and adaptive threshold three-frame difference method was used to detect the moving target,and the optimal three-frame difference threshold was adaptively set by the 3 a criterion,which saved the time for manually setting the threshold and improved the efficiency of the algorithm.Then,the multi-feature fusion method was used to fuse the shape feature,global feature,local feature and scale invariant feature of the human body,namely,the aspect ratio,effective area ratio,centroid change rate,MHI-HOG feature and SURF feature of the human body.The characteristics of the MHI-HOG and SURF features are reduced by PCA,and the five features are serially combined into one feature vector.Finally,the fusion feature was input into the SVM classifier for training.The experimental results showed that the method can detect falls in real time and has a higher accuracy.2.A fall detection method based on 3D convolutional neural network was proposed.Firstly,the moving target was detected by a combination of mixed Gaussian and adaptive threshold three-frame difference method.Secondly,to extract the optical flow characteristics of the moving target,this paper adopted a new input,namely,the optical flow history image,which contained more effective information than the original image and the optical flow stacked image.Then,the optical flow history image was input into the 3D convolutional neural network to learn the characteristics of time and space.This paper constructed two 3D convolutional neural network models to perform feature training.Finally,using Softmax for behavior classification.The experimental results showed that the method reduced the training time of the model and the convergence was faster.The average recognition rate can reach 93.71%.3.Based on the above research work,we designed and implemented a fall detection system.The system was mainly implemented on the first method in this paper. |
