Behavior Recognition Based On Convolution Neural Network

Recent years,the convolution neural network has been gradually popularized in the field of computer vision.Behavior recognition algorithm based on convolutional neural network has achieved good results,because of the problems which are not easily to solve,such as great changes in appearance,occlusion,non-rigid motion,change of scale,change of perspective,subtle movements are not obvious or the similarity between background and target,behavior recognition algorithm based on convolutional neural network there is still much room for improvement in detection performance and detection speed.Aim at the study of single behavior recognition,we put forward three kinds of behavior recognition algorithm based on convolutional neural network,four channel behavior recognition algorithm based on dilated residual network DRN,a weighted behavior recognition algorithm based on two databases,a network fused by Faster R-CNN and DRN.The main research work and innovation are as follows:(1)Two kinds of recognition algorithm based on DRN model convert each image of the color database to grayscale image,then pass through high pass filter and normalized,with high frequency grayscale picture formed.The four-channel dataset consists of color images and its high frequency grayscale picture corresponding,which is used for behavior recognition based on DRN.We use the high-frequency grayscale dataset made up of all high-frequency grayscale images and the original color dataset to predict behaviors in a weighted way,which is the weighted behavior recognition algorithm based on two databases.The experimental results show that two behavior recognition algorithms based DRN have obtained 76.1% and 76.7% in m AP respectively,which are better than other algorithms.(2)Aim at the problem that the above two algorithms can not locate the foreground target,we replace general convolutional layer in Faster R-CNN with dilated convolution residual block proposed in DRN,with a fusion model formated.and put forward two kinds of improved fusion model based on the fusion model: add a Batch Normalization layer in front of each layer;Use three layers of dilated convolution block instead of partial two layers of residual block.The experimental results show that the fusion model has a higher mAP than the other algorithms and two algorithms mentioned above.Among them,the fusion model contain three layers of dilated convolution residuals is the highest,which is 78.9%.