AI-Driven Autonomous Surveillance Drone

Overview

A custom quadcopter platform engineered for security and surveillance workloads that would traditionally require an operator in the loop and a ground processing pipeline. The airframe carries an on-board AI companion computer running YOLOv7 for real-time object detection, license-plate reading and face recognition directly against the camera feed.

A PX4-based flight controller handles autonomous mission execution and obstacle avoidance, while a MAVLink telemetry link streams position, video and system state back to a Mission-Planner-based ground control station.

The design goal was operational autonomy: the drone plans and flies its own missions, makes perception decisions on-board without a cloud round-trip, and continues to operate meaningfully in GPS-degraded conditions using visual/inertial cues. The platform was built end-to-end — mechanical frame, avionics integration, perception stack and ground software — and validated in both simulation and outdoor field trials.

System Architecture

Hardware

Software & AI

The perception stack runs a custom-trained YOLOv7 model on the on-board Raspberry Pi, targeting surveillance-relevant classes (people, vehicles, bags, intruders) via transfer learning from COCO weights. Training-time augmentation covered rotation, brightness normalisation and motion blur to reflect real airborne footage. The deployed model is optimised with TensorRT to hit the frame-rate budget required for live detection during flight.

A SORT (Simple Online and Realtime Tracking) layer maintains identity across frames for smoother tracking of moving subjects.

Autonomous flight combines A*-based waypoint planning against the mission map, fuzzy-logic obstacle avoidance fed by IR and ultrasonic sensor data, and an EKF that fuses IMU, barometer and GPS into a robust pose estimate. The GCS was implemented as Mission Planner alongside a custom Python + OpenCV dashboard that overlays live YOLOv7 detections onto the video feed.

Key Engineering Work

• End-to-end airframe engineering — SolidWorks-based design, CNC-machined carbon-fibre parts, and full mechanical / avionics integration into a flight-ready platform
• Custom YOLOv7 training pipeline (transfer learning from COCO, targeted augmentation) for surveillance classes, and TensorRT-optimised deployment on the Raspberry Pi 5
• SORT-based multi-object tracker layered on top of the detector to hold identity across frames during airborne motion
• Autonomous mission stack on PX4 with A* planning, fuzzy-logic obstacle avoidance and EKF-based state estimation fusing IMU, GPS and barometer
• Custom GCS dashboard combining Mission Planner with a Python/OpenCV overlay that unifies live detections, telemetry and mission state in one operator view
• Simulation-first validation strategy: MATLAB dynamic modelling, Simulink + PX4 software-in-the-loop, and hardware-in-the-loop testing before every outdoor flight

Results & Validation

Simulation results (MATLAB + Simulink + PX4 SITL + HIL) tracked field results within expected margins, giving high confidence that the on-board control loop matched the modelled dynamics prior to outdoor flight.

Technical Stack