APPLIED COMPUTER VISION & EMBEDDED SYSTEMS

Aervion — Smart Aerial Navigation System

AI-powered autonomous drone navigation, visual target identification, click-to-lock tracking, and decoupled asynchronous flight control.

1. Project Overview & Context

Aervion is an applied computer vision and drone flight control engineering project built to explore autonomous vision-guided following on lightweight quadcopter platforms (DJI Tello SDK).

The system integrates live video feed streaming, real-time face detection, target tracking, click-to-lock target selection, voice control triggers, and an interactive browser-based dashboard built with Streamlit.

2. The Core Engineering Challenge

In initial prototypes, running deep neural face detection on the main loop caused severe frame-rate drops. Whenever image preprocessing or inference took longer than 80ms, the flight-control loop missed vital command cycles.

Root Cause:

Quadcopter flight stability relies on a continuous high-frequency heartbeat and Proportional-Integral-Derivative (PID) correction loops. Coupling expensive image inference directly to the flight thread led to control latency, erratic yaw/pitch corrections, and safety disconnects.

The primary engineering objective was to achieve smooth, jitter-free target following without sacrificing flight safety or control frequency.

3. The Decoupled Engineering Architecture

We redesigned the system into an asynchronous, multi-threaded pipeline that isolates heavy vision processing from flight telemetry:

Thread A: Vision & Inference Worker
  • • Ingests 720p UDP video stream from drone camera
  • • Runs OpenCV face detection & bounding-box calculation
  • • Computes target error offset (dx, dy, target area)
  • • Updates shared thread-safe state queue non-blockingly
Thread B: Flight Control Loop (50Hz)
  • • Reads latest coordinates from thread-safe state
  • • Evaluates PID algorithm for Yaw, Pitch, and Throttle
  • • Maintains continuous SDK heartbeat & emergency failsafes
  • • Transmits RC velocity commands via UDP sockets

Result: Even if image inference fluctuates due to lighting conditions or complex scenes, the flight stabilization thread continues uninterrupted, ensuring smooth, predictable quadcopter tracking.

4. Verified System Capabilities

Face Detection Haar Cascades & DNN face detectors with dynamic confidence filtering.
Target Lock & Follow Bounding box tracking with automated PID yaw and distance estimation.
Safe Following Margins Target bounding box area thresholds maintain safe stand-off distances.
Manual & Voice Modes Keyboard manual overrides and speech recognition command integration.
Streamlit Telemetry HUD Interactive web cockpit displaying battery level, flight speed, and live feed.
Emergency Failsafes Instant land, hover, and connection watchdog triggers on signal interruption.

5. Technology Stack

CORE LANGUAGE
Python 3.10+
COMPUTER VISION
OpenCV (cv2)
DRONE PROTOCOL
DJI-Tello SDK / UDP
DASHBOARD UI
Streamlit

Explore More Engineering Projects

Review our other architecture prototypes and case studies across CRM automation, AI copilot systems, and data pipelines.

← Back to Case Studies Start a Project →