Towards a Benchmark for Trajectory Prediction of Autonomous Vehicles

George Daoud, Mohamed El-Darieby

Abstract: The technology stack of connected and autonomous vehicles (CAV) consists of sensing, perception, motion prediction, and motion planning Layers. With much success, the sensing and perception layers have been developed. Recently, R&D activities on the prediction of vehicles trajectory have been attracting a lot of attention as it has the potential to increase safety for road users. Trajectory prediction is a significantly more difficult task because it involves capturing historical patterns of vehicle movements that requires an understanding and analysis of unstructured spatial and temporal data at the same time. Datasets that are used for this research are typically incomplete or not generic enough. Machine learning prediction models are developed in a bit of an ad hoc manner that they use various evaluation metrics. In this paper, we discuss the issues of such datasets, models, and evaluation metrics. We also present the requirements and initial high-level design of a benchmarking software framework that allows model users to search for and select already developed models, contributed by model developers, that process data collected by dataset contributors, and evaluated by the proposed framework. Further design and development of the proposed framework will ensue.