A Deep Learning-Based Approach for Vehicle Motion Prediction in Autonomous Driving

Abstract

The prediction of a safe collision-free trajectory is probably the most important factor preventing the full adoption of autonomous vehicles in a public road. Despite recent advancements in motion prediction utilizing machine learning approaches for autonomous driving, the field is still in its early stages and necessitates further development of more effective methods to accurately estimate the future states of surrounding agents. This study introduces a novel deep learning approach for detecting the future trajectory of surrounding vehicles using a high-resolution semantic map and aerial imagery. Our proposed approach leverages integrated spatial and temporal learning to predict future motion. Specifically, it incorporates recurrent and convolutional neural networks in a novel mechanism to capture both visual and temporal features. We assess the efficacy of our proposed approach on the the Lyft Level 5 prediction dataset and achieve a comparable performance on various motion prediction metrics. Our approach has promising applications in autonomous driving and shows the potential to aid in the creation of safer and more efficient transportation systems.