Assessing the Adaptability of Deep Learning-Based Motion Prediction Models for Diverse Traffic Scenarios

Inhalt:
A smooth and comfortable steering of autonomous vehicles (AVs) in traffic is challenging. It is necessary to anticipate the evolution of traffic in order to plan a smooth ego trajectory. Despite the development of various approaches to capture traffic behavior, most models and datasets focus primarily on right-hand traffic scenarios. In this study, we investigate the transferability of a model trained on right-hand traffic data to left-hand traffic, focusing on the model’s ability to navigate left turns, right turns, and straight driving scenarios. We studied a simple LSTM-based model that does not process semantic as well as map information in contrast to VectorNet, a prominent graph-based motion prediction model. We make use of the Argoverse and NuScenes datasets. To account for the lack of left-hand traffic scenarios in the Argoverse dataset, a coordinate transformation was applied. The NuScenes dataset contains actual left-hand traffic data for analysis and comparison. In a further step, the models were fine-tuned on the left-hand data. Our results indicate that models trained for right-hand traffic can only be further adapted to left-hand traffic situations to a moderate extent through fine-tuning. Surprisingly, VectorNet generally does not outperform the simple LSTM-based model in terms of ADE. These results suggest that explicit consideration of the spatial relationships between road elements and agents may improve prediction performance only in certain cases, but may not necessarily provide a significant advantage over simpler models, depending on the dataset, although VectorNet responds much better to fine-tuning with left-hand data.