Accurate Natural Trail Detection Using a Combination of a Deep Neural Network and Dynamic Programming

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Trail detection in unstructured natural environments is a significant challenge for autonomous systems.
• Existing methods struggle with the variability and lack of clear features in natural terrains.
• Deep neural networks (DNNs) show promise for image processing but require refinement for complex tasks like trail following.

Purpose of the Study:

• To develop and evaluate a robust vision sensor-based system for detecting and following trails in challenging natural environments.
• To combine the pattern recognition capabilities of DNNs with the optimization power of dynamic programming for improved trail segmentation.
• To demonstrate the system's effectiveness using real-world data from a head-mounted vision system.

Main Methods:

• A patch-based deep neural network (DNN) was trained to classify image patches as 'trail' or 'non-trail'.
• The DNN was adapted into a fully convolutional architecture to generate trail segmentation maps for various image sizes.
• Dynamic programming was employed to refine the DNN's output, identifying an optimal trail path from the segmentation map.

Main Results:

• The combined DNN and dynamic programming approach achieved accurate trail detection in complex natural environments.
• Experimental results validated the system's performance on real-world trail datasets.
• The method demonstrated robustness despite the inherent ambiguity of trail features in unstructured terrains.

Conclusions:

• The proposed vision-based system effectively addresses the challenge of trail detection and following in unstructured natural environments.
• The integration of deep neural networks and dynamic programming offers a powerful solution for robust autonomous navigation.
• This research contributes a significant advancement in robotic perception for outdoor applications.