Urban Driving Based On Condition Imitation Learning And Multi-period Information Fusion

In recent years,autonomous driving has been a hot topic,and it is a kind of unmanned driving technology that is realized by intelligent pilots of computer systems.It plays an important role in solving traffic congestion and reducing traffic accidents,and has a significant impact on future travel,urban planning,and infrastructure construction.Among them,urban roads are intricate and complicated,and there is a large number of people and vehicles.Traffic accidents and traffic congestion occur most frequently.This is the main scenario for the application of autonomous driving technology in the future,and it is also a huge challenge for the current autonomous driving technology.In the traditional vision-based driving scheme,the entire driving strategy is divided and packaged into different rule modules,which performs well in simple driving scenarios such as highways,but it is difficult to be competent in the complex and changing urban driving environment.The algorithm based on deep learning,by directly learning driving strategies from the data of human driving demonstration,that is,directly outputting control information in an end-to-end manner,has shown good results in complex driving environments.Although the traditional conditional imitation learning algorithm solves problems such as intersection ambiguity,it lacks interpretability.The network’s feature extraction and expression capabilities are limited.There are still problems in local planning and detailed implementation,such as slow straight roads and car bodies.Swing,drive into the reverse lane,etc.In response to the above problems,we propose to use a deep residual network architecture and add a dual attention module to learn driving skills closer to humans.First,by using a deeper residual network architecture,the ability to extract detailed features of the network is further improved.Compared with the simple network framework used by traditional imitation learning,deep networks have more powerful learning capabilities to meet the data requirements of multi-task learning.Then,by introducing a dual attention module,the global context long-range dependence of the image in the spatial and feature dimensions is adaptively integrated to improve the network expression ability.Secondly,in order to make full use of the multi-period attribute information that comes with the camera image itself,the network architecture was redesigned to extract and integrate features,which increased the interpretability of the model timing information.Finally,based on the conditional imitation learning of multi-period attribute information fusion,a corresponding automatic driving decisionmaking system is designed.Combining regularization and end-to-end deep learning,the input,decision-making and control are divided into three independent modules with no coupling between the modules,which avoids the complicated and bloated and selfcontradictory of the traditional decision-making system.In this thesis,our method is tested on the CARLA driving simulator.The experimental results show that compared with the benchmark algorithms,it achieves better driving effect.Deeper feature extraction and multi-period information fusion can effectively improve the driving ability and driving completion of the agent.