ForaNav is an insect-inspired, online target-oriented navigation system designed for Micro Air Vehicles (MAVs) operating in tree plantations. The system integrates real-time tree detection with bio-inspired navigation strategies to enable efficient and autonomous UAV flight in cluttered environments.
Our paper is available now [PDF]. The demo video is available on Youtube now [Video].
This repository contains two main components:
Tree detection: To achieve real-time oil palm tree detection on resource-restricted MAVs, we use HOG features and an SVM classifier. Our method further improves detection by distinguishing oil palm trees from similar palm species and background objects with star-shaped features.
Insect-inspired navigation: Our navigation strategy draws inspiration from insect foraging behavior, particularly the reliance on local visual cues and their body-centered frame of reference for position estimation and efficient movement.
If you use this code in an academic context, please cite our work:
Weijie Kuang, Hann Woei Ho, Ye Zhou and Shahrel Azmin Suandi, ForaNav: Insect-inspired Online Target-oriented Navigation for MAVs in Tree Plantations.
(The code will be released upon publication.)
The Tree detection folder contains the implementation of the proposed tree detection approach. It includes:
- SVM Training Code: Scripts for training the proposed Histogram of Oriented Gradients (HOG)-based SVM model.
- Lightweight Deep Learning Training Code: Implementation of deep learning-based tree detection for comparison, including MobileNetV2 and EfficientNet.
- Pre-trained SVM Model: The trained SVM model for oil palm tree detection.
- HOG-based Detection Algorithm: The proposed tree detection method using the trained SVM model.
- Oil Palm Tree Dataset: The dataset used for training and evaluating the models.
To distinguish different background targets from the template oil palm tree image I-T, a combination of HOG and color-based features is utilized. For example, image I-1, which represents backgrounds such as grass and shrubs, has a similar color distribution to I-T but can be effectively distinguished using HOG feature visualization. In contrast, image I-2, including targets like buildings, shares a similar HOG structure with I-T, yet can be separated based on differences in the Hue and Saturation (H&S) histogram. Image I-3, representing targets such as banana trees and exhibiting similarity in both color and structure, is differentiated from I-T by analyzing the variance in the HOG directional histogram. This is because the needle-like leaflets of oil palm trees show greater directional variety compared to other palm-like species.
The Insect-inspired navigation module implements our insect-inspired target-oriented navigation strategy. This approach enables MAVs to navigate efficiently in tree plantations using local visual cues rather than relying on GPS or pre-mapped environments. Key features include:
- Bio-Inspired Strategy: The system mimics insect foraging behavior, leveraging egocentric visual navigation and view memory.
- Local Visual Guidance: The MAVs utilize a body-centered frame of reference to estimate position and adjust movement dynamically.
- Real-Time Adaptation: The navigation algorithm continuously updates flight paths based on detected trees, ensuring efficient tree tracking.
- Lightweight & Efficient: Designed for resource-constrained MAVs, the method provides real-time performance without the need for heavy computational resources.
This navigation module works in tandem with the tree detection system to enable autonomous flight through plantation environments.
The MAV platform used in this research consists of the following components:
