A Real-Time FPGA-Based Edge Computing Approach for Pedestrian Trajectory Prediction

This study presents an FPGA-based edge computing solution for real-time pedestrian trajectory prediction in autonomous driving applications. The system integrates three key components: pedestrian detection using the YOLOv5m model, multi-object tracking with ByteTrack, and trajectory prediction. To address the computational constraints of edge devices, we employ model compression techniques including channel pruning and quantization. Our implementation on the Xilinx ZCU104 platform demonstrates significant improvements in computational efficiency: the detection model achieves an 85% reduction in parameters while maintaining robust performance with only a 0.5% decrease in mAP, and increases inference speed from 19.03 FPS to 32.18 FPS. The complete system achieves real-time trajectory prediction performance suitable for autonomous driving applications. This work advances the field of sustainable edge computing by effectively balancing computational efficiency, prediction accuracy, and real-time performance requirements for safety-critical autonomous driving systems.