Vision-Based Autonomous Drone in Indoor
GPS-Denied Environment
1 Introduction and Objective
The primary objective of this minor project is to develop and implement a robust autonomous
navigation system for an Unmanned Aerial Vehicle (UAV) operating in indoor, GPS-denied en-
vironments. Traditional drone navigation relies heavily on Global Navigation Satellite Systems
(GNSS), which fail to provide reliable signals inside buildings or underground facilities.
This project aims to solve the localization and path-planning problem by utilizing Vision-
Based State Estimation and Simultaneous Localization and Mapping (SLAM). The system will
enable the drone to build a real-time map of its surroundings, estimate its own pose (position
and orientation), and navigate autonomously from a start point to a target destination while
dynamically avoiding static and dynamic obstacles.
2 Proposed Methodology
To achieve autonomous indoor flight, the system will rely on a fusion of onboard sensors and
computer vision algorithms. The methodology is broken down into three core operational
phases:
2.1 Vision-Based State Estimation
Vision-based state estimation is a fundamental technique used for localization when GPS sig-
nals are unavailable. We will use a depth-sensing camera to extract and track visual features
(corners, edges) across consecutive frames. By calculating the movement of these features, we
can estimate the drone’s motion.
Because camera data can be computationally heavy and sometimes slow, we will perform Sensor
Fusion by integrating camera data with high-frequency Inertial Measurement Unit (IMU) data.
We will implement a Stereo Multi-State Constraint Kalman Filter (S-MSCKF) to combine
stereo vision and IMU data for precise motion tracking.
2.2 Mapping and Perception
As the drone moves, it needs to understand its environment to avoid collisions. We will utilize
an Intel RealSense depth camera for 3D mapping and terrain analysis. The raw data from
the depth sensors and an optical flow sensor (PMW3901) will be fused to obtain a coherent
representation of the environment. The mapping module will generate a real-time Occupancy
Grid Map, identifying which areas are free space and which contain obstacles.
1
2.3 Trajectory Planning and Obstacle Avoidance
Once the map is generated, the drone must determine the most efficient path to its goal. We will
implement the A* (A-star) algorithm, a graph traversal and pathfinding algorithm that finds
the shortest path between nodes while avoiding obstacles. To ensure the drone’s movements
are physically viable and do not cause instability, the calculated path will be smoothed using
Bezier curves, which generate continuous curves to avoid sharp corners in the trajectory.
3 System Architecture
The system architecture is divided into low-level flight control and high-level computation:
• Flight Controller: A Pixhawk or STM32-based flight controller will handle low-level
flight stabilization, motor control, and raw IMU data acquisition.
• Companion Computer: A powerful microcomputer, such as a Raspberry Pi 4 or Nvidia
Jetson Nano, will act as the brain of the drone. It will process the heavy SLAM algorithms,
run the AI-driven obstacle avoidance, and send velocity/position commands to the flight
controller via the MAVLink protocol.
• Software Stack: The Robot Operating System (ROS) will be used to manage sensor
nodes, integrate algorithms, and bridge communication between the companion computer
and the flight controller via MAVROS. Feature extraction and visual odometry will be
handled using the OpenCV library.
4 Required Components and Cost Estimate
The following table outlines the hardware required to build the indoor autonomous testbed.
The estimates are based on standard market rates for prototyping components.
[Link] Component / Hardware Details Quantity Estimated Cost (INR)
1 F450 Carbon Fiber Frame 1 2,700
2 1000 KV BLDC Motors & 30 AMP ESCs 4 sets 4,500
3 Flight Controller (Pixhawk/STM32) 1 8,500
4 Companion Computer (Raspberry Pi 4) 1 6,000
5 Intel RealSense D435i Depth Camera 1 35,000
6 Optical Flow Sensor (PMW3901) 1 2,500
7 3S LiPo Battery (5500mAh) & Power Module 1 5,000
8 Propellers, Shock Absorbers, Fasteners 1 set 1,500
Total Estimated Cost ≈ 65,700
Table 1: Estimated Bill of Materials for Indoor Drone Prototype
5 Project Plan
• System Design & Procurement: Finalize hardware architecture, order necessary com-
ponents, and set up the ROS environment on local simulation machines.
• Hardware Assembly & Base Setup: Assemble the drone frame, mount the flight
controller and companion computer, and tune PID controllers for stable manual hovering.
2
• Sensor Integration: Interface the Intel RealSense and optical flow sensors with ROS.
Implement the Vision-Based State Estimation and S-MSCKF algorithms using initial
flight data.
• Mapping & Path Planning: Develop and integrate the local occupancy grid mapping
and A* trajectory planning algorithms onto the companion computer.
• Indoor Testing & Optimization: Conduct controlled flight tests in indoor, GPS-
denied environments. Fine-tune obstacle avoidance parameters and evaluate autonomous
navigation accuracy.
• Project Wrap-up: Compile flight data, assess performance metrics, and complete the
final project documentation and presentation.