基於深度學習之視覺式即時室內健身輔助系統

Abstract

近年來人們的健康意識抬頭，越來越多人開始重視規律健身習慣的養成，而進行健身運動中的肌力鍛鍊時，記錄健身的內容是避免受傷及追求進步的重要途徑。因此本研究提出一種基於深度學習之視覺式即時室內健身輔助系統，能夠辨識健身動作並進行重複次數的計算，目的在於協助使用者在不需接觸式設備的情況下更加便利的自動記錄健身內容。視覺式即時室內健身輔助系統可分為三個模組，分別為健身動作辨識模組、動作重複次數計算模組及推理校正模組。本研究使用較適用於行動裝置的改良版Temporal Shift Module來進行健身動作辨識，並利用健身動作辨識神經網路所擷取的特徵圖進行動作重複次數計算，藉由訊號過濾、訊號選擇及峰值過濾演算法篩選出合適的特徵值變化訊號與峰值。最後將動作辨識結果利用短期推理分數和長期校正分數推理校正，使其具備穩定性同時保有使用者對於動作更換的敏感度。當使用者更換動作時，重複次數計算模組的結果將被重置，並輸出整合後的結果。本研究進行實驗的肌力鍛鍊動作共有25種，分別為一般深蹲、相撲深蹲、分腿蹲、前跨步蹲、後跨步蹲、羅馬尼亞硬舉、臀推、橋式、單腳橋式、跪姿伏地挺身、伏地挺身、臥推、仰臥飛鳥、俯身划船、反向飛鳥、肩推、前平舉、側平舉、肱二頭肌彎舉、肱三頭肌伸展、俄羅斯轉體、仰臥起坐、捲腹、交叉捲腹及仰臥抬腿。實驗結果顯示使用CVIU Fitness 25 Dataset在動作辨識的Top 1準確率為90.8%。動作重複次數計算之平均絕對誤差MAEn為8.36%，次數相對誤差MREc為2.55%，在總次數5004次中計次總差異量為183次。系統執行速率約為13 FPS。In recent years, people's health awareness has risen, and more and more people have begun to pay attention to the development of regular fitness habits. Recording the content of fitness when performing muscle strength exercises is an important way to avoid injuries and pursue progress. Therefore, this study proposed a vision-based real-time indoor fitness assistance system based on deep learning, which can identify fitness movements and calculate the number of repetitions, with the purpose of assisting users to recordfitness content more conveniently and automatically without contact device.The vision-based real-time indoor fitness assistance system can be divided into three modules, namely fitness action recognition module, action repetition counting module, and inference correction module. In this study, an improved version of the Temporal Shift Module, which is more suitable for mobile devices, is used for fitness action recognition, and the feature maps captured by the fitness action recognition neural network is used to calculate the number of action repetitions. Filter out suitable feature value change signals and peaks through signal filtering(SF), signal choice(SC), and peak filtering(PF). Finally, the result of the action recognition module is corrected by short-term inference score and long-term correction score, which make the system stable and maintain its sensitivity to the user's changing actions. The results of the repetition counting module will be reset when the user changes the action, and the integrated result is output. A total of 25 muscle strength exercises were experimented in this study, including Squat, Sumo Squat, Split Squat, Lunge, Reverse Lunge, Romanian Deadlift, Hip Thrust, Bridge, Single Leg Bridge, Knee Push Ups, Push Ups, Bench Press, Chest Fly, Bent-Over Row, Reverse Fly, Shoulder Press, Front Raise, Lateral Raise, Biceps Curl, Triceps Extension,Russian Twists, Sit Ups, Crunch, Bicycle, and Leg Raise. The experimental results show that the Top 1 accuracy of fitness action recognition using CVIU Fitness 25 Dataset is 90.8%. The Mean Absolute Error in number of miscounted videos (MAEn) of action repetition counting was 8.36%, the Mean Relative Error of count (MREc) was 2.55%, and the total difference was 183 times in the total number of 5004 times. The system execution rate is about 13 FPS.