Open Access

ARTICLE

A Lightweight UAV Visual Obstacle Avoidance Algorithm Based on Improved YOLOv8

by Zongdong Du^1,2, Xuefeng Feng³, Feng Li³, Qinglong Xian³, Zhenhong Jia^1,2,*

1 College of Computer Science and Technology, Xinjiang University, Urumqi, 830046, China
2 Signal Detection and Processing Autonomous Region Key Laboratory, Xinjiang University, Urumqi, 830046, China
3 Xinjiang Uygur Autonomous Region Research Institute of Measurement and Testing, Urumqi, 830000, China

* Corresponding Author: Zhenhong Jia. Email:

Computers, Materials & Continua 2024, 81(2), 2607-2627. https://doi.org/10.32604/cmc.2024.056616

Received 26 July 2024; Accepted 25 September 2024; Issue published 18 November 2024

Abstract

The importance of unmanned aerial vehicle (UAV) obstacle avoidance algorithms lies in their ability to ensure flight safety and collision avoidance, thereby protecting people and property. We propose UAD-YOLOv8, a lightweight YOLOv8-based obstacle detection algorithm optimized for UAV obstacle avoidance. The algorithm enhances the detection capability for small and irregular obstacles by removing the P5 feature layer and introducing deformable convolution v2 (DCNv2) to optimize the cross stage partial bottleneck with 2 convolutions and fusion (C2f) module. Additionally, it reduces the model’s parameter count and computational load by constructing the unite ghost and depth-wise separable convolution (UGDConv) series of lightweight convolutions and a lightweight detection head. Based on this, we designed a visual obstacle avoidance algorithm that can improve the obstacle avoidance performance of UAVs in different environments. In particular, we propose an adaptive distance detection algorithm based on obstacle attributes to solve the ranging problem for multiple types and irregular obstacles to further enhance the UAV’s obstacle avoidance capability. To verify the effectiveness of the algorithm, the UAV obstacle detection (UAD) dataset was created. The experimental results show that UAD-YOLOv8 improves mAP50 by 3.4% and reduces GFLOPs by 34.5% compared to YOLOv8n while reducing the number of parameters by 77.4% and the model size by 73%. These improvements significantly enhance the UAV’s obstacle avoidance performance in complex environments, demonstrating its wide range of applications.

---

Keywords

---