DAT: Deep Learning-Based Acceleration-Aware Trajectory Forecasting.

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Autonomous driving (AD) systems require robust safety features, with object detection and trajectory forecasting being critical for collision prevention.
• Current AD systems face challenges in accurately predicting the future movements of surrounding vehicles and pedestrians.

Purpose of the Study:

• To introduce the Deep learning-based Acceleration-aware Trajectory forecasting (DAT) model for enhanced object detection and trajectory forecasting in AD systems.
• To leverage raw sensor measurements and acceleration data for more accurate prediction of agent motion.

Main Methods:

• Developed an end-to-end deep learning model (DAT) that processes sequential sensor data for object detection and trajectory forecasting.
• Introduced a novel forecasting module utilizing acceleration data and a method for estimating ground-truth acceleration.
• Integrated an object detector that predicts acceleration attributes and a new trajectory forecasting method.

Main Results:

• DAT was trained and evaluated on the NuScenes dataset, demonstrating significant improvements over state-of-the-art methods.
• Achieved up to a 2x improvement in forecasting accuracy, particularly for objects with complex linear and nonlinear motion patterns.
• Empirical validation confirmed the effectiveness of incorporating acceleration data into predictive models.

Conclusions:

• The DAT model represents a substantial advancement in autonomous driving safety by improving trajectory forecasting accuracy.
• Incorporating acceleration data is crucial for developing more reliable and safer autonomous driving systems.
• The proposed methods for acceleration estimation and trajectory forecasting offer a robust solution for real-world AD applications.