Research On Human Action Detection Based On Deep Learning

Human action detection is an important research field in computer vision and pattern recognition.It has a wide range of applications in many fields such as motion analysis,intelligent video surveillance,human-computer interaction,civil and military applications.Human action detection is to identify action in video images by computer vision and computer graphics and image processing.Traditional human action testing needs to extract behavioral features by hand,extracting effective features is difficult and the recognition accuracy is low.Deep learning method can enable multi-level computing models to learn multiple levels of abstract data representation,and has made significant progress in speech recognition,visual object recognition and other fields.Therefore,it is effective and worthy of exploration to study human action detection by means of deep learning.In this paper,we mainly study the method of human action detection by deep learning convolutional neural network.Convolutional neural network,compared with fully connected neural network,can greatly reduce the number of parameters to be trained by local receptive field and weight sharing,so that the training speed is faster,the efficiency is higher,and the effective features can be extracted into human action.In the human behavior detection,this paper focuses on the Faster R-CNN algorithm,aiming at improving the algorithm's shortcomings,and realizes the algorithm from data collection to final output detection information.Firstly,the principle and training process of the convolutional neural network are studied,and then select Faster R-CNN algorithm,also the principle of this algorithm is elaborated in detail.The dataset of human action is made by using the standard dataset in Pascal VOC2012,and alternating training method is selected according to the characteristics of the algorithm,and the specific algorithm flow is determined.And choose the deep learning open source framework Caffe to perform this algorithm.Secondly,the two pre-training models were used to initialize and extract features of human action dataset,respectively.Faster R-CNN algorithm was used to experiment.By comparing the results of the two models through experiments,it is found that the deeper the network of the deep learning model,the more the structure of the model is and the better the recognition effect on various behaviors in human action will be.Finally,some improvements are made to the algorithm based on the above conclusions.In the improvement,we mainly consider adding a deeper model ResNet,to the batchnormalization algorithm and the online hard-example mining algorithm,which make the deep network easier to train.The experimental results show that the improved mAP is more than80%,and the experimental results show that the improved algorithm has the characteristics of high recognition accuracy.