TrajAngleNet: Transformer-Based Trajectory Prediction Through Multi-task Learning with Angle Prediction

Vibha Bharilya, Neetesh Kumar

Abstract: Accurate trajectory prediction is crucial for the safe and autonomous navigation of self-driving cars. By analyzing sensor data and historical information, it forecasts the movements of surrounding entities. Nevertheless, the accuracy of trajectory prediction models in autonomous systems is significantly compromised when training alone with trajectory supervision, particularly for complex maneuvers such as lane changes and sharp turns. A multi-task trajectory prediction system called TrajAngleNet is presented to address the aforementioned issue. By incorporating angle prediction as an auxiliary task, the model gains the directional context and orientation of agents, thereby improving the accuracy of future trajectory predictions in dynamic environments. Further, we adopt a classification approach to address the challenges posed by the wrap-around behaviour of angles. This method mitigates numerical instability in regression by discretizing angles into bins and focusing on probability distributions rather than exact values. Moreover, the representation of angles in the input using sine and cosine values proves effective in managing periodicity and discontinuities. The TrajAngleNet model has been evaluated on the widely used and recent Argoverse 2 dataset and outperforms state-of-the-art models by a margin of 1.83–35.58%, 1.74–34.68%, 6.25–64.77% on the minADE\(_1\), minFDE\(_1\), and MR\(_6\) respectively.

External IDs:dblp:conf/iconip/BharilyaK24