Prediction and navigation for heterogeneous agents

Abstract

Autonomous driving has attracted much attention of both academia and industry in recent years. It is a very significant task because it has the potential to change people's day-to-day life and make artificial intelligence to better serve human life. At the same time, it is a very difficult task because it requires high-quality devices and mature algorithms to guarantee the safety and efficiency. Especially in urban environment, challenges arise sharply due to the density and complexity of heterogeneous obstacles. Current autonomous driving system contains four sequential modules, perception, prediction, navigation, and control. In this thesis, we mainly focus on prediction and navigation problems in challenging scenarios.

The agent's navigation will be safer if it knows what will happen in next few seconds. In urban traffic scenarios, autonomous vehicles must make responsible predictions in relation to surrounding traffic-agents (vehicles, bicycles, pedestrians, etc.) to safely and efficiently move to the destination. A critical but difficult task is to explore the movement patterns of heterogeneous traffic-agents and predict their future trajectories accurately to help the autonomous vehicle make reasonable navigation decision. To solve this problem, we propose a long short-term memory-based (LSTM-based) realtime traffic prediction algorithm, TrafficPredict. Our approach uses an instance layer to learn instances' movements and interactions and has a category layer to learn the similarities of instances belonging to the same type to refine the prediction. We collected trajectory datasets in a large city consisting of varying conditions and traffic densities and evaluate the performance of TrafficPredict on the new dataset.

In multi-agent navigation, a fundamental task is to make sure the agent has no collision with others. For complex scenarios where obstacles have different shapes and sizes, we present a novel algorithm for reciprocal collision avoidance between heterogeneous agents. We propose a novel CTMAT representation based on medial axis transform to compute a tight fitting bounding shape for each agent. Based on the reciprocal velocity obstacle formulation, we reduce the problem to solving a low-dimensional linear programming. The overall runtime performance of our algorithm is comparable to prior multi-agent collision avoidance methods that use circular or elliptical agents. Our approach is less conservative and results in fewer false collisions. As we all know, in traffic scenarios of real world, different vehicles have different kinematics and dynamics. Then, we extend the general navigation algorithm to traffic scenarios by taking into account kinematic and dynamic constraints of various traffic-agents. We evaluate the performance of our simulation algorithm in real world dense traffic scenarios and highlight the benefits over prior schemes.