# Developing a Compressed Object Detection Model based on YOLOv4 for Deployment on Embedded GPU Platform of Autonomous System

**Authors:** Issac Sim, Ju-Hyung Lim, Young-Wan Jang, JiHwan You, SeonTaek Oh, Young-Keun Kim  
**arXiv:** [2108.00392](https://arxiv.org/abs/2108.00392)  
**Date:** 2021-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.00392v1/

## Key points

- YOffleNet is a new object detection model designed for real-time and safe driving applications on autonomous systems.
- Existing CNN-based models are accurate but require high-performance GPUs, making them unsuitable for embedded systems with limited memory space.
- Lightweight detection models have low accuracy for safe driving applications.
- YOffleNet is based on the YOLOv4 backbone network architecture but replaces the high-calculation-load CSP DenseNet with lighter modules from ShuffleNet.
- YOffleNet achieves a 4.7 times higher compression ratio compared to YOLOv4-s.
- Experiments using the KITTI dataset show that YOffleNet achieves real-time performance with as fast as 46 FPS on an embedded GPU system (NVIDIA Jetson AGX Xavier).
- Despite the high compression ratio, the accuracy of YOffleNet is only slightly reduced to 85.8% mAP, which is just 2.6% lower than YOLOv4-s.
- YOffleNet offers a promising solution for overcoming memory limitations in embedded systems without compromising performance or safety requirements in autonomous vehicles.

## Summary

This paper presents a new object detection model called YOffleNet, which is designed to be compressed at a high ratio while minimizing accuracy loss for real-time and safe driving applications on an autonomous system. The existing CNN-based object detection models are accurate but require high-performance GPUs, making them unsuitable for embedded systems with limited memory space. While there are lightweight detection models available, their accuracy is too low for safe driving applications. To address these challenges, the proposed YOffleNet model is based on the YOLOv4 backbone network architecture. However, instead of using the high-calculation-load CSP DenseNet, the model replaces it with lighter modules from ShuffleNet. This significant compression allows YOffleNet to achieve a 4.7 times higher compression ratio compared to YOLOv4-s. The experiments conducted using the KITTI dataset demonstrate that YOffleNet can achieve real-time performance with as fast as 46 frames per second (FPS) on an embedded GPU system (NVIDIA Jetson AGX Xavier). Despite the high compression ratio, the accuracy of YOffleNet is only slightly reduced to 85.8% mean average precision (mAP), which is just 2.6% lower than YOLOv4-s. Overall, this study highlights the potential of deploying the proposed YOffleNet network on embedded systems in autonomous vehicles for real-time and accurate object detection applications. By compressing the network while preserving accuracy, this model offers a promising solution for overcoming memory limitations in embedded systems without compromising performance or safety requirements.