Deep Learning Methods for Vehicle Trajectory Prediction: A Survey

Trajectory prediction in autonomous vehicles deals with the prediction of future states of other vehicles/traffic participants to avoid possible collisions and decide the best possible manager given the current state of the surrounding. As the commercial use of self-driving cars is increasing rapidly, the need for reliable and efficient trajectory prediction in autonomous vehicles has become eminent now more than ever. While techniques or approaches that rely on principles and laws of physics and traditional machine learning-based methods have laid the groundwork for trajectory prediction research, they have primarily been shown to be effective only in uncomplicated driving scenarios. Following the increasing computing power of machines, the said task has seen an increase in popularity of DL techniques, which have demonstrated high levels of reliability. Driven by this increased popularity of deep learning approaches, we provide a systematic review of the popular deep learning methods used for vehicle trajectory prediction. Firstly, we discuss the problem formulation. Then we classify the different methods based on 3 categories—the type of output they generate, whether or not they take social context into account, and the different types of deep learning techniques they use. Next, we compare the performances of some popular methods on a common dataset and finally, we also discuss potential research gaps and future directions.