Time- and Resource-Efficient Time-to-Collision Forecasting for Indoor Pedestrian Obstacles Avoidance.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Real-time object detection and depth estimation are crucial for autonomous systems.
• Lightweight solutions are needed for embedded devices in navigation systems.
• Obstacle detection and collision prediction are challenging for small devices like drones.

Purpose of the Study:

• To develop a novel, lightweight, and time-efficient vision-based solution for predicting Time-to-Collision (TTC).
• To create a navigation system module for visually impaired pedestrians using smartglasses.
• To enhance obstacle detection and collision prediction capabilities for embedded systems.

Main Methods:

• A two-module system: static data extractor (CNN for obstacle position/distance) and dynamic data extractor (fully connected network for TTC prediction).
• Utilizes a monocular video camera embedded in a smartglasses device.
• Employs supervised learning for the Time-to-Collision network to adapt to new sceneries and obstacles.

Main Results:

• The proposed system demonstrates a novel approach to real-time Time-to-Collision prediction.
• The convolutional neural network effectively extracts static obstacle data.
• The fully connected network successfully predicts TTC by processing dynamic obstacle data over frames.

Conclusions:

• The developed system offers a promising solution for real-time collision prediction in embedded navigation systems.
• The Time-to-Collision network shows adaptability to diverse environments and obstacles through supervised learning.
• This research contributes to safer navigation for visually impaired individuals by leveraging smartglasses technology.