Researches On Dynamic Obstacle Avoidance Method Based On Deep Learning For Intelligent Vehicle

In recent years,with the rapid development of artificial intelligence technology,research in the field of intelligent vehicle has also rapidly emerged.In the research of intelligent vehicle,not only the support of cutting-edge technology in hardware systems,but also the technological breakthroughs in vehicle intelligent driving software systems play a vital role.The entire intelligent driving software system includes modules for environment awareness,decision planning and control.In the decision-making planning module,it is extremely important to realize the dynamic obstacle avoidance for intelligent vehicles.In the human-road-vehicle intelligent transportation system,in order to realize the safe driving of the intelligent vehicle,it is necessary to ensure that the vehicle can avoid dynamic or static obstacles in real time.However,most of the current methods for implementing dynamic obstacle avoidance of intelligent vehicles are traditional rule-based methods.Based on the detection of obstacles,rules are set to allow vehicles to avoid obstacles.It is difficult to meet all the constraints.In addition,the environment of different roads where vehicles are located is complex and changeable.The behaviors of other vehicles and pedestrians also have great variability.Therefore,we take pedestrians as the research object and conduct pedestrian trajectory prediction with Deep Learning method and Deep Reinforcement Learning method for intelligent vehicle dynamic obstacle avoidance.Because human drivers can make better decision-making in terms of perception and prediction after accumulating rich experience,but human beings will be affected by physical and psychology.Therefore,considering the shortage of human drivers,we do the research on dynamic obstacle avoidance based on human-machine cooperation with Deep Reinforcement Learning.The main research results and innovations of this paper are as follows:(1)The Social-Grid LSTM method for pedestrian trajectory prediction is proposed.This method is based on a Long-Short Term Memory Network,combining the Social Pooling operation and the Grid LSTM model,in which the Social Pooling operation is to establish a relationship between the interactions between pedestrians.The reason for adopting the Grid LSTM model is mainly that compared with the general LSTM model with the same number of layers,the value of the loss function converges faster,and the stable value of the final convergence is smaller.The training and testing are carried out in the two public pedestrian trajectory data sets.The performance comparison with the LSTM method and the Social LSTM method is carried out.The experimental comparison results show that the proposed method can get smaller prediction errors.(2)The dynamic obstacle avoidance system for intelligent vehicles is designed with Deep Q-Learning method,with pedestrians as the research object.As the dynamic obstacle,the state data of the vehicle and the state data of the pedestrian as the state in the RL,and then the longitudinal and lateral control of the vehicle are designed in the actions.In the experiment,the dynamic model of the vehicle is considered to be close to the real vehicle.After online training,the vehicle can be driven at different speeds.In the process of implementing the algorithm,we designed two types of experience replay,including experience replay with positive and negative samples and negative samples only.By the experience playback of negative samples,the agent can be effectively trained with the collision situation to improve the performance of the dynamic obstacle avoidance.We set up the environment for training and testing on the Prescan software.The experimental results show that the dynamic obstacle avoidance strategy can achieve better obstacle avoidance performance.(3)A dynamic obstacle avoidance method based on DRL and human-machine coordination is proposed.It combines human-machine coordination mechanism to design a security restriction condition.During the controlling of vehicle by the human driver,the safety restriction conditions in the design model are compared by the current driving behavior and the state of the vehicle,and then in the case of a certain danger,that is,the safety restriction condition is not met,the vehicle control will be switched to the DRL agent's strategy,and then the output control action automatically controls the vehicle to avoid the collision.The simulation under different scenarios is designed on the Prescan software.The experimental results show that the human-machine cooperation with DRL agent can improve the safety performance of the vehicle.