Abnormal Behavior Recognition Based On Improved I3D Neural Network

At present,the application of computer vision in automatic analysis of video sequence images is developing rapidly,among which the most applied fields are human moving target detection and abnormal behavior recognition.In the above-mentioned sports scene,it is mainly to realize human motion analysis and action recognition under dynamic conditions,and continue to judge whether the action belongs to the category of abnormal behavior after recognizing the human behavior action.Firstly,this thesis improves the inception structure based on the original I3D,the convolution kernel is replaced with two convolution kernels according to the principle of convolution kernel replacement,and the improved model is trained on UCF101 data set,which reduces the training time about 9%Secondly,this thesis makes use of the idea of channel shuffle,the channel shuffle strategy is integrated into the improved I3D neural network,and a new 3D network model(I3D-shufflenet)is proposed.According to class activation mapping results,I3D-shufflenet is better than the original I3D network in terms of image feature extraction and image effective recognition area.Finally,the I3D-shufflenet network is applied to identify the abnormal action on the bus,and a framework of identifying the abnormal action based on the bus camera is built.First of all,this thesis collects about 50 videos of abnormal action,such as bus fighting,and extracts 10 frames per second to gather the initial data set.However,the image resolution and pixels are low.Then,the GAN network is adopted to re-generate the images.The final data set has about 15,000 pictures with high resolution,and the frames of abnormal action are labelled.The dataset is divided into training set and test set by the ratio of 5:1.The I3D-shufflenet neural network and I3D neural network are trained on bus abnormal action data set,and the accuracy rate is 69.8%and 64.5%respectively.Compared with I3D neural network,both the recognition rate and the convergent performance of I3D-shufflenet neural network are significantly improved.The proposed model is the improvement of I3D model.The validity of the proposed model is verified by experiments.Compared with other action recognition models,I3D-shufflenet neural network has obvious improvement.