Motion Focus Recognition in Fast-Moving Egocentric Video

TL;DR

The paper tackles the challenge of extracting locomotion intent from fast-moving egocentric video, where appearance-based cues are insufficient for reliable ego-motion understanding. It introduces a real-time motion focus recognition pipeline that leverages a camera-pose foundation model, a sliding-window inference scheme with incremental anchoring, and an acceleration-based projection to generate a motion saliency map, while maintaining depth $\mathbf{D}_t$ and extrinsics $\mathbf{T}_t^{w \rightarrow c}$ for downstream reasoning. Key contributions include a memory-efficient real-time inference strategy and a physically grounded motion focus mechanism that ties image-space attention to 3D motion via $a_w$ and $a_c$, producing a map $M(u,v)$. Evaluations on a winter egocentric dataset show real-time operation on consumer GPUs and motion-focused attention that aligns with the videographer’s locomotion, enabling edge deployment for sports, robotics, and safety applications.

Abstract

From Vision-Language-Action (VLA) systems to robotics, existing egocentric datasets primarily focus on action recognition tasks, while largely overlooking the inherent role of motion analysis in sports and other fast-movement scenarios. To bridge this gap, we propose a real-time motion focus recognition method that estimates the subject's locomotion intention from any egocentric video. Our approach leverages the foundation model for camera pose estimation and introduces system-level optimizations to enable efficient and scalable inference. Evaluated on a collected egocentric action dataset, our method achieves real-time performance with manageable memory consumption through a sliding batch inference strategy. This work makes motion-centric analysis practical for edge deployment and offers a complementary perspective to existing egocentric studies on sports and fast-movement activities.

Figures (5)

• Figure 1: (a) Selected frame from the dataset collected during the winter season. (i) is the raw city street image, (ii) is the motion-focus map, and (iii) is the motion-guided depth map. (b) is the example camera system for data capturing.
• Figure 2: Framework for motion focus prediction and projection.
• Figure 3: Depth alignment in the real-time inference. We applied a sliding window for consecutive inference in real time.
• Figure 4: Motion focus calculation process. We first calculate the camera center and project the acceleration vector to the $k$th frame. Then we aggregate $K$ camera centers and apply Gaussian kernels for rendering. The final results represent the camera movement trends, which aligned with the defined motion focus.
• Figure 5: Visualization of motion focus. Left: raw image with predicted motion focus map; Right: depth map guided by motion direction.