EPIC: Ego-Centric Monocular 3D Pedestrian Trajectory Estimation via Instance-Aware Center-Focused Depth Processing

Engineering notes

Key topics: autonomous driving system, autonomous driving, depth estimation, instance segmentation, object tracking, lidar, monocular camera, nuscenes. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

As autonomous driving systems become more prevalent, ensuring the safety of pedestrians is a critical challenge that must be addressed. One of the most crucial components of pedestrian safety systems is the ability to estimate pedestrian trajectories accurately in the vehicle coordinates. However, existing approaches primarily concentrate on 2D image-plane bounding-box forecasts, which are often inadequate for real-world applications. Other methods rely on expensive 3D sensors such as LiDAR, which hinders their commercialization. Some monocular camera-based approaches have explored 3D localization, but most remain limited to frame-level estimates, lacking the temporal continuity needed for reliable trajectory analysis. To overcome these limitations, this paper introduces EPIC: Ego-centric Pedestrian Trajectory Estimation via Instance-aware Center-focused Depth Processing, a cost-efficient system for estimating 3D pedestrian trajectories using only a monocular camera. This approach integrates instance segmentation, object tracking, monocular metric depth estimation, and an adaptive Center-focused depth extraction with Gaussian-weighted EMA smoothing, enabling trajectory estimation in the vehicle coordinate frame. This paper evaluates the proposed system on the nuScenes dataset, demonstrating the feasibility of monocular trajectory estimation for pedestrian safety applications.

Links and sources

• Official / arXiv page

Comments