Multi-Agents Trajectory Prediction for Autonomous Vehicles with Multi-Modal Predictions

This paper presents an advanced deep learning framework designed to improve motion prediction in autonomous driving, enhancing safety through better environmental understanding. The Crat-Pred model is enhanced, demonstrating superior performance over existing benchmarks in both qualitative and quantitative analyses, with improved training times and computational efficiencies. The proposed approach addresses a critical research gap by predicting the motion of both vehicles and pedestrians. By utilizing convolutional neural networks and replacing Long-Short-Term Memory (LSTM) networks with Temporal Convolutional Networks (TCN), significant improvements in computational speed and efficiency are achieved. The primary objective is to develop a unified prediction model for all road agents, advancing autonomous navigation. Experiments using the Argoverse 2 dataset validate the model's performance in real-world conditions, showcasing its superiority in forward motion prediction and computational efficiency.