Relation Learning for Information Reliability Representation and End-to-End Robust Navigation

View the PDF file from the paper entitled Conquering Ghosts: The relationship learning to represent the reliability of information and strong movement from the end to the end, written by Kivan Jin and other authors

PDF HTML (experimental) view

a summary:Environmental disorders, such as sensor data, different lighting conditions, difficult aging, and external litigation disorders are inevitable in real self -driving applications. Research and current tests have shown that it can severely affect the ability and performance of vehicles, and it is one of the main issues is the wrong positive detection, that is, the unknown ghost object or occurs in the wrong situation (such as a vehicle that is not present). Traditional mobility methods tend to avoid all the things discovered for safety, however, avoiding a ghost body may lead to a more dangerous position, such as a sudden break on the highway. Given the different types of disorders, it is difficult to address this problem in the cognitive aspect. The possible solution is to discover the ghost by learning the relationship between the entire scenario and developing an integrated navigation system from one side. Our basic logic is that the behavior of all vehicles in the scene is affected by their neighbors, and that regular vehicles act in a logical way, while ghost vehicles do not do so. By learning the spatial and temporal relationship between the surrounding vehicles, the reliability of the information is learned for each vehicle discovered and then a robot navigation network was developed. Unlike current works, we encourage the network to learn how to represent reliability and how to collect all information with uncertainty in itself, and thus increase efficiency and generalization. To the best in the knowledge of the authors, this paper provides the first work to use the relationship between the graph to achieve strong to one side in the presence of ghost vehicles. Simulation results on the Carla platform show the feasibility and effectiveness of the proposed method in different scenarios.