Prediction of Traffic Agents' Trajectories in Urban Scenarios Using Deep Learning

Abstract

Trajectory prediction of surrounding traffic agents is essential for autonomous vehicles to identify and alleviate potential conflicts. Deep learning-based (DL based) methods have gained attention because of their ability to process complex traffic scene contexts, like agent-to-agent and agent-to-road interaction. Extensive research has been conducted on trajectory prediction in highway scenarios, but limited studies focus on urban environments because of the complexity of interactions. This project selects a naturalistic driving dataset targeting urban intersections and trains DL-based models to predict the trajectories of high-mobility traffic agents. Traffic agents’ track data and corresponding traffic scene data, like neighbour and road layout (RL) information, are extracted as training samples from the selected dataset. A novel method based on maximal disks is proposed to extract lane centrelines, which are important RL features but not available in most datasets. A trajectory prediction backbone is proposed based on an RNN-based encoder-decoder structure. Experiments showed that particular architectural decisions, like using LSTM instead of GRU, explicit calculation of the velocity of traffic agents, and close coupling between the encoder and the decoder, can improve the prediction accuracy. Further, the influences of agent interaction and road scene mining were studied separately. A model using Convolutional Social Pooling to extract agent interaction was replicated from the literature and was improved for better performance. In road scene mining, map element features are extracted from the sparsely defined RL map, and agent-map attention is used to extract road scene context. The models are trained in multiple scenarios with diverse RLs. Experimental evaluations have shown that both agent interaction and road scene mining can improve prediction accuracy. This project treats the agent interaction and road scene mining separately, future work includes exploring the fusion of these two encodings and more efficient neural networks in arranging traffic scene elements and extracting corresponding features.