Title: AIGaitor: Enabling Private, Cloud-Free Motion Analysis for All via Edge Computing

Abstract:

While motion capture remains the benchmark for assessing human movement, its integration into clinical settings is hindered by high costs, technical hurdles, and privacy risks. AIGaitor addresses these challenges as a privacy-centric, cloud-independent motion analysis platform. It executes markerless, monocular motion-capture workflows and subsequent deep-learning evaluations entirely on consumer smartphones, leveraging on-device neural accelerators.

The design of AIGaitor was driven by a survey of 74 rehabilitation clinicians. The results indicated strong interest, with 92% expressing willingness to adopt an AI-driven gait analysis tool provided it was accurate, affordable, and user-friendly. However, significant barriers were identified: 79.7% pointed to operating costs, 68.9% to inadequate training, and 64.9% to privacy concerns.

To validate the system, we optimized and benchmarked mobile iOS implementations of existing monocular pipeline components. These included 2D and 3D pose estimation, pose optimization, skeleton-based deep-learning analysis, and a vision-language model. Our Time-Priority, end-to-end on-device pipeline processed a 10-second, 4K, 60 fps video clip in just 77 seconds on an iPhone 14. This performance matches or exceeds that of the same pipeline running on a high-end NVIDIA H200 cloud server, even when accounting for network transfer times: 94 seconds at global average mobile uplink speeds and 66 seconds on developed-world Wi-Fi. Furthermore, lightweight models like ViTPose-s enable real-time keypoint extraction, while skeleton-based action-recognition models classify gait in under a millisecond for the same video.

To our knowledge, AIGaitor represents the first monocular system to demonstrate end-to-end, on-device motion capture combined with downstream deep-learning analysis. This approach facilitates clinically viable movement analysis that is low-cost, secure, and accessible to users with standard smartphones.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC