Pedestrian Navigation

Abstract

Intelligent devices for supporting persons with vision impairment are becoming more widespread, but they are lacking behind the advancements in intelligent driver assistant system. To make a first step forward, this work discusses the integration of the risk model technology, previously used in autonomous driving and advanced driver assistance systems, into an assistance device for persons with vision impairment. The risk model computes a probabilistic collision risk given object trajectories which has previously been shown to give better indications of an object's collision potential compared to distance or time-to-contact measures in vehicle scenarios. In this work, we show that the risk model is also superior in warning persons with vision impairment about dangerous objects. Our experiments demonstrate that the warning accuracy of the risk model is 67% while both distance and time-to-contact measures reach only 51% accuracy for real-world data.

Navigation, Obstacle and AI (NOA) device

NOA device

New intelligent, camera-based devices are designed to detect and alert users of obstacles in real time. By acting as an extra set of eyes, these devices are transforming personal safety, offering enhanced independence and reducing the risk of falls and head-level injuries. One example for such a device is NOA (Navigation, Obstacle and AI.), developed by Biped Robotics. NOA is a wearable shoulder vest designed to assist the navigation of people with vision impairment. The device perceives the environment through camera sensors and runs a vision processing pipeline for object detection.

Risk Model

NOA device

In this work, we analyze to what extend state-of-the-art collision estimation methods used in autonomous driving and advanced driver assistance systems, can improve the warning accuracy of the NOA blind assist device. To this end, we integrate the risk model, a probabilistic risk estimation framework, into NOA. The object detection pipeline is mainly based on processing point cloud data generated from RGB-D sensors. Object trajectories are derived in a linear manner by comparing object positions from two consecutive time steps. The risk model then uses growing Gaussian distributions to model trajectory predictions and uncertainty.

Evaluation Setup

Evaluation Pipeline

We use two different datasets to test the improvements of the risk model. The first dataset was recorded using the NOA device in various indoor and outdoor environments (sidewalks, university campus and train stations). It includes 30k frames captured along with IMU data. The second dataset was generated using Webots, a 3D robotics simulator. It includes 10k frames and consists of very specific scenarios such as side collisions or walking towards a crowd. By using different setups in the datasets, we created 5 different experiments that include different levels of noise and errors (see Experiment 1-5).

Results

	Plain			Hysteresis			Hysteresis + JPDAF
Experiment	Risk	TTC	Distance	Risk	TTC	Distance	Risk	TTC	Distance
1 Scene Object States	86.78%	83.99%	69.08%	89.33%	91.39%	76.21%	89.33%	91.39%	76.21%
2 Annotated Simulated Data	73.00%	75.84%	68.90%	72.86%	77.24%	72.27%	80.64%	80.21%	74.08%
3 Simulated Data	66.39%	60.62%	57.59%	66.03%	64.86%	58.81%	74.36%	66.37%	65.90%
4 Annotated Camera Data	28.46%	48.88%	36.60%	59.57%	54.68%	48.91%	70.52%	63.33%	61.85%
5 Camera Data	21.43%	17.89%	28.18%	23.58%	24.89%	31.35%	67.00%	51.33%	50.90%

We compared the risk model with two baseline approaches based on distance and Time-To-Contact (TTC). The table shows the performance measured by the intersection over union (IoU) metric, averaged across all data from the five different experimental setups. We report results for the plain approaches, as well as versions enhanced with hysteresis and with both hysteresis and a Joint Probabilistic Data Association Filter (JPDAF). The risk model achieves the highest accuracy in experiments involving realistic data setups. For instance, using hysteresis + JPDAF on real-world camera data from the NOA device, the risk model has an accuracy of 67% while both distance and TTC reach only 51% accuracy.

BibTeX

For more details, please refer to our paper. If you find our work helpful, kindly cite it as follows:

@inproceedings{tourki2025,
      author = {Tourki, Amine and Prevel, Paul and Einecke, Nils and Puphal, Tim and Alahi, Alexandre},
      title = {Probabilistic Collision Risk Estimation for Pedestrian Navigation},
      booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
      year = {2025}
}