Research On Pedestrian Anomaly Behavior Recognition Algorithm Based On Multi-fiber Feature Fusion

With the continuous development of computer vision and deep learning theory,behavior recognition technology based on deep learning has gradually become one of the important research directions in the field of computer vision,and it is widely used in many fields such as intelligent monitoring and human-computer interaction.Among them,Convolutional Neural Network(CNN)has made a breakthrough in image classification tasks as a typical deep learning model compared with traditional feature extraction methods.However,the behavior recognition model applied to video still has problems such as low accuracy,computational cost and large model storage.Therefore,on the basis of deep learning,it has a wide range of research value and important research significance to identify the behavior in video effectively.Firstly,this paper introduces the background and significance of pedestrian abnormal behavior recognition technology,analyzes the research status at home and abroad,then expounds the theoretical basis of convolutional neural network and the mainstream behavior recognition method based on deep learning,analyzes the current problems and difficulties in this task,and based on the problems in pedestrian abnormal behavior recognition,this paper has carried out the following two aspects of work and innovation:(1)A 3D Multi-fiber Network Based on Depthwise Separable Convolution is proposed.In view of the large computational complexity of the 3D CNN model,this paper introduces common model compression methods of 2D CNN in the 3D Multi-fiber Network(3D MF-Net),which is excellent in current behavior recognition tasks.First of all,the Depthwise Separable Convolution is added to the network convolutional layer,and a multi-fiber unit module with a relatively high computational cost is replaced by a structure based on Depthwise Separable Convolution.Compared with the calculation cost of the original model,it can effectively reduce the computational cost.Secondly,with the idea of inserting space into the convolution kernel to make it expand in the Dilated Convolutions,the larger convolution kernels in the network are dilated to reduce the amount of calculation.By weighing the relationship between model complexity and accuracy under different width multipliers,the model achieves an effective compromise between efficiency and performance,and finally a 3D Multi-fiber Network Based on Depthwise Separable Convolution(3D DSMF-Net)is realized,the experimental results show that the model sacrifices less precision and brings about a large reduction in the amount of calculation and parameters.(2)A 3D Adaptive Convolutional Neural Network Based on Attention Mechanism is proposed.Aiming at the problem of precision loss in model compression process,this paper constructs a 3D Adaptive Convolutional Neural Network Based on Attention Mechanism(3D ADNet-AM)by combining three modules of 3D adaptive convolutional layer and multi-scale feature fusion pooling layer(3D Adaptive)to improve the validity and robustness of the model for abnormal behavior recognition.The 3D adaptive convolution layer extends the selective convolution unit in the Selective Kernel Networks(SKNet)to three dimensions,the module adjusts the size of the convolution layer receptive field adaptively according to the input information,realizes feature extraction of spatio-temporal information of two scales,and uses the Attention Mechanism(AM)to generate the attention vector as the weight when the two types of features are fused;The multi-scale feature fusion pooling layer is separately pooled by three different scales of the original feature maps,and the output features are spliced in a cascade manner to obtain a fixed-dimensional feature map.Compared with 3D MF-Net,this algorithm not only reduces the computational cost of 2.81 G,but also improves the Top-1 accuracy of 2.45%,which proves that the proposed algorithm has a significant improvement on network performance.