Standardized detection and counting of sit-ups based on CenterNet

Motion detection technology based on deep learning can adapt to the impact on recognition accuracy in a variety of complex backgrounds, but the existing algorithms are difficult to ensure real-time and accurate detection of sit-ups. Aiming at this problem, a scheme for the detection and counting of sit-ups based on the Centernet algorithm is proposed. Collect 14,000 samples for training and testing. And analyze the action characteristics of sit-ups, and establish a parameter classification model to judge whether the exercise is performed correctly. In order to improve the recognition accuracy of the Centernet algorithm, an attention mechanism is introduced to pay attention to detailed information to improve the accuracy of detection. In terms of network optimization, deep separable convolution is used to reduce the amount of network parameters. The final experimental results show that the accuracy of the scheme is 91.90%, and the precision is 92.51%. The average FPS reached 30 at the camera resolution of 480P, proving that the solution can be used for sit-up detection and counting accurately and in real time.