Research On Behavior Recognition Of Vehicle Forward Pedestrian Based On Skeleton Nodes

With the increase of the number of cars,the traffic accidents of cars and pedestrians are increasing.As the most vulnerable group of traffic participants,pedestrians are the most valued in the traffic environment.Therefore,in the environment perception part of intelligent driving,the behavior recognition technology of pedestrians is particularly important.The continuous development of deep learning technology provides an important foundation for the research of pedestrian behavior recognition.In this paper,analysis the method of pedestrian behavior recognition in front of vehicle is studied and verified by real vehicle test.The main research contents are as follows:(1)Pedestrian skeleton node detection based on VGG-19 network model.Based on the characteristics of real-time,stability and accuracy of pedestrian detection,a bottom-up pedestrian skeleton node detection algorithm is implemented by using VGG-19 network model,which can get the information of pedestrian skeleton node in complex traffic environment,and prepare for the subsequent pedestrian feature extraction.(2)Feature extraction of pedestrian behavior based on Resnet50 network.The cosine value of key joint is calculated by skeleton node information and converted into gray-scale image.On this basis,the feature of gray-scale image sequence is extracted by resnet50 network of convolution neural network.The residual learning idea of the network solves the problem of information loss and loss of traditional convolution neural network to a certain extent,reduces the learning difficulty on the premise of ensuring the integrity of information,and effectively completes the feature extraction of pedestrian behavior.The results show that the method has a high rate of feature extraction of behavior.(3)Pedestrian behavior recognition based on LSTM network.Based on the extraction of pedestrian behavior features from resnet50 model,the time series problem of feature sequence frame is analyzed by using cyclic neural network LSTM,and the final behavior classification is completed by using Soft Max classifier.Aiming at the typical pedestrian crossing behavior,the model training is carried out to realize the pedestrian behavior classification.(4)Vehicle verification of the model.In the Keras framework with Theano as the backend,use the behavior classification model to classify the behavior of pedestrians in the real scene,to realize the pedestrian behavior recognition of intelligent vehicles,and to verify the effectiveness of the model in this paper.In conclusion,this thesis combines the deep convolution network and the recurrent neural network to study the spatiotemporal correlation between the skeleton feature points and the motion intention,and realizes the pedestrian behavior recognition,which can provide the theoretical basis for the deep environment perception and adaptive decision-making technology of intelligent vehicles.