Optimization of Video Repetitive Action Counting for Efficient Inference on Edge Devices

Repetitive actions are prevalent in both natural and man-made environments, offering valuable insights into the analysis of action units and underlying phenomena. Video repetition counting task aims to predict the count and frequency of the repetitive actions. Deep learning models have been developed for this task, enabling the recognition of repetitive motions without physical contact with the moving object. However, these models often perform unnecessary operations during inference due to inefficient data pre-processing. To address this issue, we propose an optimized data frame pre-processing method that minimizes redundant operations, ensuring fast and accurate inference. Furthermore, in order to enable video repetition counting on edge devices, we employ quantization for model compression, allowing the deployment of lightweight models suitable for various applications.