Autonomous Drone
Vision-Based Autonomous Exploration on a Lightweight Quadrotor
Project Domain
Robotics & Autonomy
Execution Stage
Research Build
Timeline / Year
2026 · Active Research & Development
Repository & Code
GitHub Repository →Visual Documentation & Artifacts
Click any photo to enlargeEngineering with real-world constraints.
Autonomous Drone is a robotics research project built around a Tello / Tello Talent quadrotor platform with a clear goal: develop a vision-first autonomy system for aerial exploration that works with low-cost hardware, without depending on expensive sensor arrays like multi-camera rigs or LiDAR units.
A primary challenge of monocular vision is fundamental scale ambiguity: while monocular visual SLAM (ORB-SLAM3) computes camera motion and 3D landmark points with accurate relative proportions, a single camera cannot resolve absolute physical dimensions. The recovered trajectory is dimensionless—a 1-meter translation indoors and a 10-meter translation in an open field produce the same normalized SLAM geometry. Without resolving this scale ambiguity, true metric mapping and autonomous path planning are impossible.
To overcome this without heavy sensor payloads, the system integrates Depth Anything V2, which outputs dense metric depth in real-world units directly from monocular RGB frames. The perception pipeline actively leverages these metric depth predictions to calculate the live scale factor, resolving the scale ambiguity of the ORB-SLAM3 trajectory. Additional hardware cues (such as an external Time-of-Flight sensor or drone telemetry) are strictly optional secondary sources. The resulting metric geometry feeds into a 3D volumetric occupancy grid and a frontier-based exploration planner.
Why This Project?
The project explores how far autonomous aerial navigation can be taken using a small, lightweight drone and vision-based perception rather than depending on a large sensor stack.
A monocular camera provides rich visual information at very low hardware cost and weight, but it introduces difficult problems: monocular scale ambiguity, visual drift, depth estimation from a single viewpoint, coordinate-frame consistency, sensor synchronization across asynchronous processes, real-time processing constraints in a resource-limited embedded environment, and reliable 3D mapping from an airborne platform that is itself constantly moving.
This repository is built around solving those problems as a complete, integrated system. Every module is designed to be replaced or extended independently, so the research direction of any component can evolve without rebuilding the entire stack.
Component Deep Dives
DJI Tello Quadrotor
│ (Wi-Fi H.264 Stream)
▼
PyAV Low-Delay Decoder
│
┌─────────┴─────────┐
▼ ▼
ORB-SLAM3 Depth Anything V2
(Scale-Ambiguous) (Metric Depth)
│ │
└─────────┬─────────┘
▼
Active Scale Recovery
(Calculates Scale Factor)
▼
3D Volumetric Occupancy Grid
▼
Frontier Exploration PlannerDrone Interface — Physical Layer
The `drone_interface` layer completely isolates the physical Tello hardware from the rest of the autonomy stack. It manages the wireless connection, video acquisition with configurable resolution and bitrate, telemetry polling, command dispatch, and manual override control.
Video is decoded from the Tello H.264 stream using PyAV with low-delay buffering options so the perception pipeline always operates on the most recent frame rather than stale buffered imagery.
This isolation is intentional: any changes to the underlying quadrotor platform or its SDK do not propagate into the perception, mapping, or planning layers, keeping the system modular and future-replaceable.
Explore Full Codebase on GitHub
Includes complete source code, installation instructions, launch files, and documentation.
Key Capabilities
Monocular Visual SLAM (ORB-SLAM3)
Computes real-time 6-DOF camera pose and sparse 3D landmark maps from the drone's forward camera stream, maintaining loop closure detection and relocalization capability as a standalone ROS 2 C++ process.
Learned Dense Depth Estimation (Depth Anything V2)
Infers high-resolution dense depth maps in real-world metric units from monocular frames, providing complete surface coverage to fill the gaps between sparse SLAM feature points.
Active Metric Scale Recovery (Depth Anything V2)
Actively resolves monocular scale ambiguity by calculating a live scale factor from Depth Anything V2's real-world metric depth outputs, converting the dimensionless SLAM trajectory into physical meters (external ToF sensors remaining strictly optional).
3D Volumetric Occupancy Mapping
Maintains a probabilistic 3D voxel grid of the explored environment, incrementally updated from fused metric depth observations, enabling collision-aware path planning.
Frontier-Based Autonomous Exploration
Identifies boundaries between known free space and unmapped regions (frontiers), selects exploration targets by value, and generates collision-free trajectories for systematic area coverage.
Isolated Drone Interface Layer
Decouples all Tello-specific hardware communication — video decoding, telemetry, command dispatch — from the autonomy stack, allowing the platform to be swapped without touching perception or planning code.
ROS 2 Humble Modular Architecture
Each subsystem runs as an independent ROS 2 node communicating over typed message topics, making the architecture extensible and each component independently replaceable as the research evolves.
- [01]ROS 2 Humble
- [02]ORB-SLAM3
- [03]Depth Anything V2
- [04]Metric Scale Alignment
- [05]OctoMap
- [06]C++17
- [07]Python 3
- [08]DJI Tello
Other Selected Works
Conditional Multimodal MRI Synthesis & Brain Tumor Segmentation
A two-model system combining a ResNet U-Net for multi-class brain tumor segmentation with a conditional diffusion model that synthesizes 4-modality MRI scans from masks. Led as a two-person project and trained on TPU.
Learn more →Computer Vision & Edge AIGesture-Controlled Drone
A real-time gesture-recognition pipeline controlling a DJI Tello in flight. Built on a self-collected dataset of 5,000+ hand-landmark samples, a feature-engineered classifier, and temporal smoothing for reliable control.
Learn more →Robotics & Embedded SystemsGesture-Controlled Robotic Hand
An MPU6050-driven glove that maps hand orientation to real-time servo control on a 4-DOF robotic arm — connecting embedded sensing directly to actuation.
Learn more →