| Autonomous driving technology is one of the key transformative technologies for the future of automobiles.As a connecting part of the autonomous driving system,the prediction module performs high-level representation and understanding of low-level environmental perception information,and outputs the future state of surrounding agents.With the help of deep neural networks,autonomous vehicle can make high-precision trajectory prediction,behavioral intention and risk level of surrounding agents,and thus the correctness and solution speed of the decision-making and planning module of autonomous vehicle was improved.Currently,deep learning-based prediction methods have surpassed rule-based and machine learning methods in terms of accuracy and generalization,and it is also a hot research topic in the field of autonomous driving.However,in the face of more complex and variable traffic scenarios,the prediction method based on deep learning still faces many difficulties and challenges to be solved urgently,such as multi-source heterogeneous input processing with noise,feature extraction and semantic fusion of multi-source heterogeneous data,and advanced semantic information sharing among multiple prediction tasks.In addition,as more and more autonomous driving technologies are deployed in commercial vehicles,deep neural networks,as black box models,have received increasing attention in terms of transparency,credibility,and interpretability.How to improve the interpretability of the deep learning model and mine the intention-behavior-movement pattern of the agent is also the bottleneck that needs to be broken through in the autonomous driving prediction task.To address these issues,this paper focuses on the risk prediction,trajectory prediction and behavior prediction of autonomous vehicle.The research content of this paper mainly includes the following aspects:First,to solve the problems of lack of labels and class imbalance in risk prediction,this study proposes a cost-sensitive semi-supervised deep learning-based risk prediction method for autonomous vehicle.This study embeds the CNN-LSTM model into a semi-supervised learning framework,and uses an adaptive overbalanced cross-entropy loss function to achieve cost-sensitive learning.The experimental results show that using only 5% and 2.5%of the total data labeling,the current moment risk prediction accuracy reaches 96.63% and95.82% respectively,which is of great significance for low-cost model training and data labeling of massive unlabeled data in autonomous vehicles.In addition,the adaptive over-balanced cross-entropy loss function can adaptively adjust the cross-entropy weights of each class according to the change in the number of instances of each risk class,so that the high-risk class is always in the over-balanced state,which improves the accuracy and recall rate of the high-risk class.Second,in order to suppress the influence of noise and uncertainty in the prediction input and improve the interpretability of the proposed model,this study proposes uncertainty-suppressed prediction model of behavior and trajectory and its interpretation method.This study proposes temporal pattern attention,graph convolutional sequence encoding to suppress low-level perceptual noise,perceptual loss,and other heteroscedastic uncertainties;An adaptive weight loss function is proposed to suppress the homoscedastic uncertainty introduced by multi-task models.Through experimental analysis,the performance of the proposed method is improved by more than 20% compared with other comparison models,and has better uncertainty suppression ability.In addition,to demonstrate that temporal pattern attention has the ability to extract high-level intention-behavior patterns,a passive model interpretation method based on perturbation input is proposed in this study.Taking the lane changing behavior of vehicles in highway scenarios as an example,statistics have found that temporal pattern attention can extract three time stages corresponding to "driving intention-lane changing behavior-vehicle motion" during the lane changing process,indicating that neural networks can effectively extract advanced "intention-behavior-motion" patterns of agents.Third,in order to solve the problems of unified representation of multi-source heterogeneous inputs and organic fusion of multi-source information in complex scenes,and to actively improve the interpretability of the model based on a priori knowledge,this study proposes a multimodal behavior and trajectory prediction method under organic fusion of multi-source information.In this study,a Holistic Transformer is proposed,which improves the performance of multimodal behavior and trajectory prediction by three different attention mechanisms.Through experimental analysis,Holistic Transformer has the best comprehensive prediction performance on all comparison models.As a regularization term for trajectory prediction,lane line intention helps the agent’s multi-modal trajectory prediction,and agent behavior prediction helps the single-modal trajectory prediction.This study also shows the visualization results of lane intention weights.Taking agent behaviors such as changing lanes and turning as examples,as the behavior progresses,the agent’s attention lanes gradually changes from the current lane to the target lane.The above conclusions are similar to human intuition,which proves that lane line intention attention has high credibility and good interpretability.Finally,in order to achieve high-precision prediction of multiple tasks and improve the interpretability of deep learning models with multi-source and multi-dimensional inputs,this study proposes a prediction integration model and its multi-source and multi-dimensional fine-grained interpretation method.In this study,the Holistic Transformer is used as the backbone network,and with the additional input of risk and behavior prediction,a branch network of risk prediction and behavior prediction is established.Compared with the traditional multi-task network framework,the multi-task network framework is based on "each task network as the main,advanced semantic information sharing as the supplement",and the total-fraction approach training is more optimal.In addition,this study proposes a passive interpretation method based on bottleneck constraints,which is analyzed in the temporal,spatial,and feature domain.In terms of feature dimension,the prediction of the lateral and longitudinal behavior of the agent is mainly based on the lateral and longitudinal acceleration of the agent itself;In the spatial dimension,the prediction of the lateral and longitudinal behavior of the agent is mainly based on the prediction of the features of the target agent itself,and secondly through the prediction of the interaction information of the adjacent agents;In the temporal dimension,taking lane changing behavior as an example,there are three highly important time nodes before the agent crosses the lane line.So far,this study completes the study of autonomous vehicle prediction task based on explainable deep learning through prediction integration model and its active and passive explanation methods. |
