Intent Prediction Of Pedestrians Based On Graph Convolutional Network And LSTM

Accurate prediction of pedestrian behavior intention is one of the most importanttechnologies in the autopilot scene.In the existing methods,pedestrian intention prediction is mostly based on trajectory prediction and interaction modeling,which makes less use of the inherent information of pedestrian body.Therefore,in this paper,while using the key points information of the bone when pedestrians cross the road,we add the key information of the face,both of which are used as the input of the prediction network.Based on the graph convolutional network(GCN)and long short-term memory network(LSTM),a pedestrian crossing intention prediction network is constructed to extract spatial sequence and time series features with good prediction results.The main contents of this paper are as follows:(1)Dataset production of pedestrian crossing.The videos of pedestrians are collected by using camera in and around the school.There are two kinds of videos: pedestrian crossing and pedestrian not crossing.Cut the collected videos of pedestrians reasonably and separate the video frames.To improve the quality of video frames,we reconstruct the resolution of video frames with low resolution.(2)Extraction of key points of bone and face.In this paper,the key points of bone and face are detected on the processed video frames,so as to predict whether the pedestrian will cross the road according to the extracted key points location information.The extracted key points of pedestrians are made into a data set in CSV format,which is convenient to input into the prediction network.In this paper,open source library Openpose is selected to extract the key points to ensure the accuracy and robustness of the detection.There are 95 key points,including 25 key points of bone and 70 key points of face.(3)Pedestrian crossing intention prediction based on the graph convolutional network(GCN)and long short-term memory network(LSTM).The processed CSV data is input into the prediction network based on the graph convolution network and the long short-term memory network.The key points information of video frames are extracted in time and space.Find the internal relationship between the pedestrian intention and the location of the key points by training network,so that the prediction network can judge whether the pedestrian will cross the road according to the change of the key points location of the pedestrian.The changes of key points of bone directly reflect the walking rules of pedestrians.The changes of key points of face can reflect people’s intention from the side and predict people’sbehavior.Therefore,we use the dual information that contains key points of bone and face as input of network for training.Combined with the ability to process spatial sequence information of graph convolution network and the ability to process time sequence information of long short-term memory network,the network can learn to predict whether the pedestrian will cross the road according to the key points information,and predict the pedestrian intention more effectively.