| In the field of intelligent driving,accurate identification of pedestrians’ intention to cross the street can avoid imminent danger in advance,and can reduce invalid waiting time,which is conducive to improving driving efficiency.In the existing researches on pedestrian crossing intent recognition,most of them use the extracted pedestrian pose features for trajectory prediction,but these methods have a single data type,are easily disturbed by environmental factors,and run slowly.Therefore,this thesis combines pedestrian detection and human key point detection in the pedestrian feature recognition stage to obtain the location information and key point coordinate information of pedestrians crossing the street,and then constructs an LSTM network model combined with multi-head attention mechanism to recognize pedestrians crossing the street to improve the accuracy and security of the detection model.The main research contents of the thesis are as follows:This thesis proposes a pedestrian crossing intent recognition method combining pedestrian detection and human key point detection.Pedestrian detection can obtain the location information of pedestrians in each frame of images,so as to determine whether pedestrians have the intention to cross the street.Human key points contain rich pedestrian action features,and the intention of pedestrians can be identified by analyzing the actions of pedestrians.By combining pedestrian detection and human key point detection,the input data types of the intent recognition model are enriched,and the accuracy of the overall method is further improved.Aiming at the problems of high missed detection rate of small targets and difficult to realize embedded real-time detection in the process of complex environment detection,this thesis adds an auxiliary residual network module and a scale to form a threescale prediction according to the basic structure of yolov4 tiny target detection network.The accuracy of the trained algorithm is 72.32%,and the FPS reaches 78 frames/s.On the other hand,the Openpose algorithm is used in this thesis to detect human key points,and identify whether pedestrians have the intention of crossing the street according to the detected coordinate information of the key points.Finally,the data of pedestrian key points and pedestrian detection data are jointly used for classification network model training and testing.A LSTM neural network model combined with multi-head attention mechanism is constructed to identify pedestrians’ crossing intentions.After the data is processed,it is input into the improved LSTM network model for training,so that the model can find the regularity between the pedestrian’s intention to cross the street and the input information.The multihead attention mechanism can give weights to different information in the operation,so that the network can learn pedestrian information samples more comprehensively and prevent data overfitting.Finally,train the model by using the mixed data and test it,compare it with the current popular algorithms,and design a real vehicle experiment to verify the accuracy and practicability of the model.The experimental results show that the accuracy rate of the improved LSTM neural network is increased by 2.5%,and the FPS reaches 18 frames/s.On the basis of the existing pedestrian detection and human key point detection technology,it can further realize the recognition of whether the pedestrian in front has the intention of crossing the street,and it is beneficial to improve driving efficiency and avoid danger in advance. |
