Research On Human Abnormal Behavior Detection Algorithm Based On Deep Learning

With the rapid development of computer vision and deep learning,a large number of scholars at home and abroad are interested in the detection of abnormal human behavior under video surveillance.And with the rapid increase in the amount of video data,how to quickly and accurately identify abnormal behaviors has become a research hotspot in this field.The detection process of abnormal human behavior consists of video preprocessing,extraction of human motion information and training of detection models,and classification of normal and abnormal behaviors.For the above process,the specific work done in this paper is summarized as follows:(1)The KNN non-parametric kernel density estimation algorithm is a commonly used method for moving target detection.It is modeled according to sample attributes and uses probability density functions to estimate data points,thereby completing the extraction of foreground images.However,when KNN nonparametric kernel density estimation algorithm is used to detect moving targets,there will be problems such as shadow extraction of the target,incomplete background separation,and unclear target contour.In view of the shortcomings and problems of the above algorithm,this paper proposes an improved KNN non-parametric kernel density estimation algorithm.This algorithm uses the K-nearest neighbor non-parameter density algorithm to calculate the probability density of pixels in the image,extract the moving target,at the same time,the Gaussian filter function is used to perform high-pass filtering on the sampled video frames in the frequency domain through the positive and negative changes of the two-dimensional discrete Fourier,and the resulting image is then merged with the foreground image,so as to extract a more visual foreground image.The experimental results show that the method proposed in this paper can get the foreground image with obvious shadow elimination effect,complete background separation and clear target.(2)In order to obtain better recognition accuracy in large data sets for human abnormal behavior detection tasks,this paper proposes the idea of using KNN non-parametric kernel density estimation algorithm combined with C3D model to detect human abnormal behavior.First,the KNN non-parametric kernel density estimation algorithm is used to process the original video to obtain the background-removed image sequence,then these image sequences are input into the C3D model for training,and finally the input behaviors are classified and detected.Experimental results show that the method proposed in this paper not only has less running time but also higher recognition accuracy.(3)Aiming at the problem that the traditional 3D convolutional neural network algorithm cannot fully utilize the multi-level convolutional features of the network,a new network model combining C3D network and fast connection is proposed for abnormal behavior detection.This algorithm adds three shortcut connections between the pooling layers in the C3D network.The shortcut connection uses a 1×1×1 convolution kernel to adjust the number of channels.At the same time,in order to prevent the neural network from overfitting,a Dropout mechanism is added after each fully connected layer.Through experiments on a large data set,the results prove that this method can make full use of multi-level convolution features,enhance the transmission of information flow in the network,and the detection accuracy is improved by 2.05%compared to C3D networks.At the same time,the method we proposed is compared with other mainstream classification models to verify the reliability of this method.