Improved YOLOv8 Based Small Object Detection for Intelligent Robotic Arm in Complex Environments

Abstract

Small objects in complex environments may be very close to the background, which brings great challenges to computer vision based intelligent robotic arms. To address these problems, an improved YOLOv8 based small object detection method is proposed in this paper. Firstly, the YOLOv8 prediction heads are reconstructed to improve the prediction performance of small objects. An adaptive multi-scale convolutional attention module (MSCA) is introduced in the backbone which leverages multi-scale features to apply spatial attention with element multiplication. Part of convolutional modules is replaced with GhostNet blocks. These GhostNet blocks generate numerous ghost feature maps with cost-effective linear transformations based on intrinsic feature maps. Finally, the coordinate regression loss function is also optimized with Wasserstein divergence. Experimental results show that the improved YOLOv8 achieves a 0.093 improvement in mean average precision (mAP) compared to vanilla YOLOv8. For practical deployment applications, a sophisticated spatially-aware intelligent robotic arm is proposed by combining the improved YOLOv8 and laser ranging. The practical application results demonstrate that the proposed method has high feasibility and stability.