Improved YOLOv5 badminton detection algorithm and embedded implementation
Liyang Cao, Zili Li · 2022
Aiming at the requirements of real-time badminton classification detection with high accuracy and lightweight under the limited resources of embedded platform, a badminton classification detection method based on improved YOLOv5 is proposed. (1) The Ghost Net network structure is adopted to compress the overall architecture of YOLOv5's backbone network layer, reduce the complexity and computation of the network, and realize the lightweight design of the algorithm; (2) The attention mechanism CBAM is introduced into C3 module to enhance the feature perception of badminton target; (3) Use a smoother Mish function to improve the Swish activation function of YOLOv5, so that feature information can flow in a deeper network layer. The experimental results show that the proposed method obtained batter performance than original YOLO V5. The detection frame rate is 55 frames/second, which shows that the algorithm is effective and meets the requirements of real-time high-precision detection.