Probabilistic Pedestrian Forecasts from a Handheld Phone: World-Frame Heat Maps, Visual-Inertial Height Drift, and Evaluation without Ground Truth

TL;DR

This paper presents a system for probabilistic pedestrian forecasting using a handheld phone with a monocular camera and visual-inertial odometry. The system detects and tracks people, forecasting their future positions as probability maps. It evaluates the forecaster without ground truth by scoring against the tracker's own later measurements.

Why It Matters

The proposed system has practical significance in applications such as autonomous vehicles, robotics, and smart cities, where predicting pedestrian movements is crucial for safety and efficiency. It can also be used in mobile devices to provide users with real-time information about their surroundings. This technology can improve decision-making and navigation in dynamic environments.

Key Findings

• The system achieves lower negative log-likelihood (NLL) scores than a baseline constant-velocity Gaussian model
• Vertical VIO drift and user's own changes of level can silently rescale monocular ground positions
• The forecaster's performance is evaluated without ground truth by scoring against the tracker's own later measurements