Research On Pedestrian Trajectory Prediction Algorithm Based On Deep Learning

With the continuous development of the intelligent automobile industry,pedestrians and vehicles are the main participants in the road environment.Accurate prediction of pedestrian walking trajectory from the perspective of vehicle has become a major research hotspot.However,there are three problems in predicting the pedestrian walking trajectory from the vehicle perspective : First,pedestrians who are prone to small targets from the vehicle perspective carry less useful feature information;secondly,pedestrian trajectory prediction is to obtain pedestrian position information through target detection algorithm,and treat pedestrians as static obstacles,ignoring the problems of strong pedestrian initiative,walking direction and walking trajectory flexibility.The third is to ignore the problem of interaction information between other pedestrians around the smart car and the target pedestrian.The target pedestrian will be affected by the walking direction and walking speed of other pedestrians around the smart car.Aiming at the above problems,this paper improves the prediction of pedestrian future trajectory from the perspective of vehicle.The research contents are as follows:Aiming at the problem that the target pedestrian is small and the feature information of the joint point is less,which cannot accurately detect the pedestrian joint point information in the vehicle perspective,the Open Pose network is improved.By increasing the resolution of the input image and increasing the image scaling ratio,the feature information of the pedestrian joint point in the vehicle perspective is more effectively captured,and the accuracy of the pedestrian joint point detection is improved.However,increasing the resolution of the input image leads to an increase in the amount of calculation of the network model.It is proposed to further improve the network structure,change the size of the convolution kernel,and use dilated convolution and depthwise separable convolution instead of standard convolution to reduce the parameters and calculations of the Open Pose network model.The experimental results show that the improved network improves the accuracy of pedestrian joint detection by 6%,reduces the number of parameters and calculation of the improved network model by 69% and 39% respectively,and improves the speed of pedestrian joint detection.Aiming at the problem that the social generative adversarial network ignores the flexibility of pedestrian walking direction and trajectory and the influence of other pedestrians around on the target pedestrian in the same scene when predicting the future walking trajectory of pedestrians in the vehicle perspective,an attention mechanism is added to the encoder in the social generative adversarial network for pedestrian trajectory prediction,focusing on the important information of the relative distance and motion direction between the target pedestrian and other pedestrians around.The weight distribution of the target pedestrian joint point information obtained by the improved Open Pose network is used to determine the main joint point of the target pedestrian and use its data as the pedestrian historical trajectory sequence,and the useful information of other pedestrians around the target pedestrian selected by the attention mechanism is used as the input of the improved social generative adversarial network generator to generate the future walking trajectory of the target pedestrian.The experimental results show that when predicting the 3.2 second pedestrian walking trajectory,the average deviation error of the improved network is reduced by 46% compared with the social generative adversarial network,and the final deviation error is reduced by 44%.The improved network model predicts that the pedestrian ’s future walking trajectory is closer to the real pedestrian ’s future walking trajectory.