ABSTRACT This project presents ARC (Adaptive Real-time Capture), a complete markerless motion capture and character animation system that achieves real-time skeletal animation of a 3D character in Unreal Engine 5 using a single commodity webcam. The core problem addressed is the longstanding cost and complexity barrier in professional motion capture: traditional systems require expensive marker suits, calibrated multi-camera rigs, and controlled environments, making them inaccessible to independent developers, small studios, and academic researchers. ARC overcomes these constraints through a carefully integrated five-stage processing pipeline. In Stage 1, video frames are acquired from a standard USB webcam (640×480, 60 fps) and preprocessed with region-of-interest cropping and contrast normalization. Stage 2 deploys MediaPipe BlazePose at complexity=2 (the HEAVY model), which produces 33 anatomical landmarks with direct sub-pixel 2D coordinates and a monocular depth estimate (z), eliminating the need for a second camera entirely. Stage 3 applies a per-joint One-Euro Filter bank comprising 99 independent adaptive filter instances — one per spatial coordinate of each landmark — which reduces per-joint jitter by 43.6% relative to unfiltered output while introducing no perceptible lag during fast motion, outperforming both simple exponential-weighted moving average (EWMA) and fixed-cutoff Butterworth approaches. Stage 4 addresses body-part occlusion through a custom quadratic temporal predictor with a 12-frame prediction horizon, backed by skeleton-graph interpolation from visible neighbouring joints when occlusion persists beyond the prediction window. This approach achieves 91.2% occlusion recovery across 15,897 test frames with no second camera. Stage 5 applies a FABRIK inverse kinematics solver over four kinematic chains (both arms, both legs) with per-joint anatomical Euler-angle limits derived from biomechanical literature — preventing the hyperextension artefacts that plague unconstrained bone-transformation methods. The processed pose data is streamed live to Unreal Engine 5 over UDP at 0.29 ms mean latency using a custom JSON protocol, driving a rigged Mannequin skeleton via a companion C++ UActorComponent plugin. Across 15,897 frames of benchmarking, ARC achieves a mean throughput of 28.7 frames per second, 43.6% jitter reduction, 91.2% occlusion recovery, and an average IK correction rate of 2.3 per frame. The entire system runs on a single consumer-grade workstation with no specialized hardware beyond a standard webcam, demonstrating the viability of professional-quality markerless motion capture at zero sensor cost.
N et al. (Sun,) studied this question.