Convolutional Block Attention Mechanism with ResNet18 for Sports Action Recognition based on Video Image Key Point Detection

Sports action recognition and evaluation depending on video image key point detection contains analyzing athletes’ movements by determining critical body joint and landmark in video frames. These key points are utilized to track motion and enable accurate classification and evaluation of different sports actions. However, it is challenging due to variability in body movements, diverse action style, and occlusions across athletes. This research proposes Convolutional Block Attention Mechanism-ResNet18 to recognize and evaluate the sports action depending on video image key point detection. In ResNet18, the CBAM is included to enhance feature representation by focusing on significant spatial and channel-wise data which improves the network’s ability to capture appropriate data. Initially, the UCF-101 dataset is used to determine the proposed approach performance. The mean subtraction is applied in pre-processing to normalize input data by centering it around zero which enhances model convergence and training stability. The proposed CBAM-ResNet18 achieves a better accuracy of 98.73% using UCF-101 dataset when compared to existing methods like Convolutional Neural Network (CNN) and Bidirectional-Gated Recurrent Unit (Bi-GRU) and Attention-based two-stream deep Neural network (A2SN).