Research On Imitation Learning And Interpretability For Autonomous Driving

With the development of the smart car industry and artificial intelligence technology,more and more Internet companies and traditional car companies have invested in research in the field of autonomous driving.As a significant innovation task,autonomous driving has also received more and more attention in the academic circle.The intelligent driving task in the urban scene is the most challenging,because driving in the town requires high-frequency interaction with the frequently changing environment.At the same time,due to problems such as urban architecture occlusion and ambient light,in the absence of high-precision maps,it is difficult for the vehicle itself to obtain all the information of other objects in the current scene.This type of incomplete information task modeling is very difficult.This dissertation is mainly to study the autonomous driving decision-making method under the condition of incomplete information.Aiming at the problem of poor generalization ability of traditional imitation learning,a curriculum learning method of offline multi-task scenarios is proposed,so that the agent can be enhanced through learning of different tasks.Meanwhile,this dissertation also adds the semantic segmentation as auxiliary tasks to improve the model’s generalization capabilities.Furthermore,an enhanced imitation learning algorithm for online data aggregation is introduced to modify the trajectory prediction of the model in the online scene through the interaction of the online scene,thereby improving the decision-making ability of the model.At the same time,the safety of autonomous driving is also a problem that companies and users have been paying close attention to.This dissertation also focuses on this problem.The decision-making process of the autonomous driving model is modeled as a process of perception and decision-making.Perception is an important part in urban driving scenes.Pedestrians,vehicles,traffic lights and other objects are detected in the urban scenes,the importance of these scene objects for action decision-making is sorted,and the most important items of the decision-making is selected,and then the actual spatial position information is merged,and finally the driving action and Interpretation is given by the model.This method can not only improve the transparency of the intelligent driving model,but also can observe the scene objects that specifically affect the decision-making after the prediction,and improve the trust of the driver and the user in the model.Finally,the classic autonomous driving simulation platform CARLA and the BDDOIA dataset are used to verify the algorithm proposed in this dissertation.Experiments are used to prove the performance of the proposed model and the effectiveness of each module.The optimal performance is achieved on both Co RL 2017 and No Crash test tasks.Compared with the previous method,the success rate is improved effectively.At the same time,this dissertation analyzes the interpretability of the decision model.Through the action prediction task and the visualization of the model output,it shows that the introduction of interpretability helps the model to understand the action decision,and the introduction of interpretable text can effectively improve the accuracy of model action prediction.