An improved algorithm for prediction of vehicle trajectories using short-term goal-driven network

Abstract

Background

Prediction of vehicle trajectories is a crucial task for automated driving systems to accurately take movement actions according to the dynamic traffic environment, especially the future positions of the surrounding vehicles. Predicting how road users will behave in the future is one of the most critical and complex challenges in autonomous driving. Different data types must be combined to accomplish this task using deep learning, such as sensor readings and maps. After that, this data is used to predict a range of possible future outcomes. Existing goal-driven approaches predict the final goal and use it to complete the trajectory; this requires accurate goal prediction, which is challenging. Short-Term Goal Network (STG) addresses this challenge using multiple short-term goals instead of a single final goal.

Results

The proposed STG network is evaluated on the Argoverse motion forecasting dataset, and the results show significantly better performance than other goal-driven approaches. STG demonstrated a substantial improvement of over 6% in average displacement error and more than 8% in final displacement error.

Conclusion

This article presents an improved algorithm for predicting vehicle trajectories using short-term goals. The proposed STG algorithm is based on long short-term memory (LSTM) and attention mechanism for predicting trajectories. This work verifies that predicting more than one goal along the trajectory improves the accuracy of the predicted goal and the whole trajectory accordingly.