Quantizing YOLOv5 for Real-Time Vehicle Detection

Abstract

Autonomous driving has received much attention in the last decade as a key component of intelligent transportation, and vehicle detection serves as a fundamental task for autonomous driving. Although recent learning-based methods have achieved great advances in terms of accuracy, these methods are usually computationally expensive and cannot be deployed on resource-limited devices. To address this limitation, in this paper, we aim to develop an efficient yet highly accurate vehicle detection method by tailoring a network quantization method. Due to the small objects and diverse scenarios in the real world, we introduce learnable scale parameters for network quantization to achieve flexible adaption. Ablation study is conducted to demonstrate the effectiveness of our method. In addition, extensive experiments are conducted on the UA-DETRAC and KITTI datasets for performance evaluation. Compared to previous generic network quantization methods, our method produces significantly better performance. Moreover, our quantized network surpasses previous lightweight vehicle detection methods in terms of both accuracy and efficiency, establishing a new baseline for vehicle detection.

Keywords

• Object detection
• lightweight modeling
• model quantization