FPGA-Based Vehicle Detection and Tracking Accelerator.

Area of Science:

• Computer Vision
• Embedded Systems
• Artificial Intelligence

Background:

• Multiobject detection and tracking algorithms face challenges with high computational complexity and power consumption on edge devices.
• Existing methods struggle to achieve high throughput and low power for real-time vehicle detection in smart transportation.

Purpose of the Study:

• To design and implement a low-power, low-latency, high-precision vehicle detector for edge devices.
• To address computational complexity and model size issues in CNN-based vehicle detection and tracking.

Main Methods:

• Utilized YOLOv3 and YOLOv3-tiny Convolutional Neural Networks (CNNs) with the Deepsort algorithm on a Field Programmable Gate Array (FPGA).
• Applied structured pruning and 16-bit fixed-point quantization to compress model size and reduce memory footprint.
• Developed a reidentification (RE-ID) dataset and trained an appearance feature extraction network for improved tracking.
• Implemented hardware optimization techniques including pipelining, memory multiplexing, and Winograd algorithms.

Main Results:

• Achieved significant model size reduction: 85.7% for YOLOv3 and 98.2% for YOLOv3-tiny.
• Demonstrated high-precision vehicle detection with low power consumption and low latency.
• Reached a peak throughput of 168.72 frames per second (fps) for 6-way parallel video stream detection.

Conclusions:

• The developed FPGA-based system offers an efficient solution for real-time vehicle detection and tracking in smart transportation.
• Model compression and hardware optimization techniques enable high performance on resource-constrained edge devices.
• The system meets the demands for high throughput and accuracy in intelligent traffic management systems.