Visual Odometry for GPS-Denied Navigation - A Hackathon Project

Earlier this month, I participated in the "Rebooting State Capacity" hackathon, an event focused on leveraging technology to enhance public services. I chose to tackle a challenge within the Defence & Security track: enabling flight navigation in environments where GPS is unavailable or unreliable. This is a critical capability gap affecting infrastructure inspection, disaster response, and defence operations.

The Problem: Navigating Blind

Many vital operations occur where GPS signals fail – inside tunnels, within collapsed buildings, or in electronically contested zones. Relying solely on GPS limits the effectiveness and increases the cost and risk associated with deploying autonomous systems like drones in these critical scenarios. How can a drone navigate when it can't rely on satellite signals?

Approach: Visual Odometry (VO)

Visual Odometry (VO) is a computer vision technique that estimates the motion of the drone/ camera by analyzing the movement of visual features in consecutive video frames.

The Hackathon Project: A Proof-of-Concept

Over the intense 24-hour hackathon period, my goal was to build a foundational proof-of-concept demonstrating VO.

• Objective: Process a video feed from a drone's perspective and estimate its 3D trajectory without GPS input.
• Tools: I primarily used Python with the OpenCV library for computer vision tasks and Matplotlib for visualization.
• Process: The core pipeline involved:
  1. Reading video frames.
  2. Detecting salient features (like corners using ORB).
  3. Tracking these features between consecutive frames.
  4. Estimating the relative camera rotation and translation based on feature movement.
  5. Accumulating these relative movements to reconstruct the overall 3D path.
• Outcome: The script successfully processed video input and generated a 3D plot visualizing the estimated flight path.

Reality Check: Limitations and Next Steps

It's important to be clear: this hackathon project was a basic implementation of VO. Techniques like this inherently suffer from drift accumulation over time and are sensitive to factors like poor lighting, lack of texture, and rapid motion. The resulting path had noticeable error.

Achieving the robustness needed for real-world autonomous drones requires more advanced techniques, primarily:

• Visual-Inertial Odometry (VIO): Fusing camera data with IMU (Inertial Measurement Unit) sensor data significantly improves accuracy, robustness, and resolves scale ambiguity.
• Visual SLAM (Simultaneous Localization and Mapping): Building a map of the environment while navigating allows the system to recognize previously seen areas (loop closure) and drastically reduce long-term drift.

These sophisticated systems are what power state-of-the-art autonomous navigation in companies like Delian Alliance and others working in the defence tech space.

Learning and Moving Forward

This hackathon was a valuable experience in rapidly applying computer vision concepts to a challenging real-world problem. It reinforced my interest in autonomous systems and provided practical insights into the complexities of visual navigation – topics highly relevant to my MSc studies in Machine Learning for Visual Data Analytics. While a simple demonstration, it highlights the potential for vision-based techniques to enhance operational capabilities where traditional methods fall short.