基於攝影機的自由重量訓練追蹤

Abstract

在運動中利用自我監控(Self-Monitoring)的機制，紀錄運動過程來量化運動成效，可以提供訓練者反饋同時增強訓練者對運動效果的信心。而重量訓練(Weight Training )是一種抵抗自身外部重量的阻力訓練，需要根據自身需求瞭解訓練目標，規劃訓練內容並執行。因此，在訓練過程中紀錄下訓練動作、重量、次數、組數和訓練/休息時間五項關鍵資訊，可以幫助訓練者評估訓練品質、衡量進步幅度以及追蹤長期訓練計畫。本研究利用電腦視覺技術，提出非接觸式的重量訓練追蹤方法。透過攝影機拍攝訓練者與訓練設備，將影像利用人體姿態估計結合物件偵測與影像分割技術，獲取人體動作與訓練設備的基礎資訊。接著，配合動作辨識模型，根據訓練者實際的自由重量訓練模式，自動追蹤動作、次數、組數、重量與訓練/休息時間五項重量訓練關鍵資訊。本研究共收集 17 位訓練者分別執行三個自由重量訓練動作的實際訓練影像，並由三個視角同時拍攝，實驗資料集共 153 部影片。針對追蹤方法進行驗證評估，內容包括五項紀錄項目。實驗結果顯示，在完整拍攝人體動作與訓練設備的多視角攝影條件下， 本研究提出的方法能準確追蹤 17 位訓練者於不同視角的訓練動作與執行組數，平均準確率可達 100% ; 此外，次數追蹤於各視角之平均F1-Score可達 0.98 ; 重量追蹤則於不同視角之準確率達 96% ; 訓練/休息時間追蹤能在 8 秒誤差容忍情況下，平均準確率達 100%， 2-6 秒誤差容忍情況下，各視角平均準確率為 93% 。綜合以上實驗結果支持所提出追蹤方法，能有效追蹤五項重量訓練內容並記錄。

In exercise, utilizing the mechanism of self-monitoring to record the exercise process can help quantify exercise efficacy and provide feedback to enhance trainers' confidence in the effects of exercise. Weight training is a form of resistance training that requires understanding training objectives, planning training content, and execution based on individual needs. Therefore, recording five key pieces of information - training movements, weight, repetitions, sets, and training/resting times – during the training process can assist trainers in evaluating training quality, gauging progress, and tracking long-term training plans. This study proposes a non-contact weight training tracking method using computer vision techniques. By filming trainers and training equipment with cameras, and utilizing human pose estimation combined with object detection and image segmentation techniques, fundamental information about human movements and training equipment is obtained from the footage. Subsequently, by incorporating action recognition models and based on actual free weight training patterns of trainers, five key pieces of weight training information – movements, repetitions, sets, weight, and training/resting times – are automatically tracked.This study collected actual training footage of 17 trainers performing three free weight training movements, filmed simultaneously from three angles, totaling 153 videos in the experimental dataset. The tracking methods were validated and evaluated for the five recorded items. The results showed that under the condition of complete multi-angle footage capturing human movements and training equipment, the proposed method can accurately track the training movements and sets of 17 trainers from different angles, with an average accuracy up to 100%; additionally, the average F1-Score for tracking repetitions can reach over 0.98; the accuracy for tracking weight can reach over 96% from different angles; tracking of training/resting times can achieve 100% accuracy under a tolerance of 8 seconds in timing error, and an average accuracy of 93% under 2-6 seconds of tolerance. In summary, the experimental results support the proposed tracking method for effectively tracking and recording five key aspects of weight training.