Loader Bucket Working Angle Identification Method Based on YOLOv5s and EMA Attention Mechanism

Abstract

In response to the issues of low recognition efficiency and large errors encountered in the process of identifying the working angle of the bucket during current automated loader construction operations, a method based on YOLOv5s and the EMA attention mechanism for loader bucket working angle identification is proposed. Initially, a small target detection head, utilizing YOLOv5s, was designed to enhance sensitivity towards target recognition. The EMA attention mechanism was introduced to increase the recognition rate of the target area and the positioning accuracy of the target frame, effectively differentiating the background area from the target area. The Focal-EIOU Loss function was added to address the slow convergence speed of YOLOv5. Subsequently, Depth Separable Convolution was employed to replace the standard convolution in the C3 module of the Backbone, improving the model’s accuracy in identifying target deformation caused by changes in the bucket angle, reducing the computational load, and enhancing the model’s operational speed. Experimental results demonstrate that the model’s mean Average Precision (mAP) value reached 99.3%, a 3.0% increase over the benchmark model YOLOv5s. The GFLOPs reached 58.5, an increase of 42, with a growth rate of 254.55%. This method effectively enhances the precision and intelligence of loader construction operations.

Keywords

• Angle identification
• YOLOv5s
• small object detection
• attention mechanism
• depth separable convolution