Real-Time exercise recognition and repetition counting from monocular gym video

This thesis presents the design, development, and evaluation of a real-time system that classifies gym exercises and counts repetitions from a single monocular camera, without wearable sensors or dedicated hardware beyond a compute backend. While pose estimation, activity recognition, and repetition counting have each been studied independently, no existing system integrates all three into a real-time pipeline that operates on commodity security cameras under realistic gym conditions. The pipeline chains three models. MediaPipe extracts 2D body landmarks from monocular RGB video, a Temporal Convolutional Network (TCN) classifies exercises across 16 categories from joint-angle time series, and RepNet counts repetitions in a class-agnostic manner. We compare five classification approaches, from scikit-learn baselines to a custom TCN. The TCN achieves 95.82% test accuracy and, unlike the baselines, produces temporally stable predictions suitable for real-time display. The complete pipeline processes each frame in 82ms on a consumer laptop and is deployed as a containerized web application. In end-to-end evaluation on 81 ground truth exercise sets, the system detects 72.8% of exercises (F1=83.7%), correctly classifies 88.1% of detected events, and counts repetitions with a mean absolute error of 2.1. A dual-camera experiment reveals that viewpoint is the primary limiting factor, with detection dropping from 70% to 40% when the camera moves from an elevated angle to a frontal position. The system currently processes one user at a time, with identity detection and multi-person tracking left as future work.