Embedded Edge-ai Framework For Autonomous Uav-based Human Detection And Remote Alerting Using Esp32-s3 Flight Control And Esp32-cam Vision

Abstract

Autonomous Unmanned Aerial Vehicles (UAVs) Equipped With On-board Artificial-intelligence Inference Are Increasingly Used For Surveillance And Monitoring Tasks That Require Rapid, Low-latency Decisions Without Continuous Human Supervision. This Paper Presents The Design And Experimental Implementation Of A Compact Embedded Edge-AI UAV Platform For Autonomous Defense-oriented Human Monitoring, Built Around A Custom-developed Quadcopter Whose ESP32-S3 Microcontroller Performs Real-time Flight Control While A Separate ESP32-CAM Module Performs On-board Visual Inference. The Flight-control Subsystem Interfaces With An Inertial Measurement Unit (MPU6050) And Pulse-width-modulated Motor Drivers To Support Stabilized Takeoff, Hover, Landing And Coordinate/waypoint-based Autonomous Navigation Over A Crazyflie-compatible Wi-Fi Command Interface. For Visual Intelligence, Two Edge Impulse Models Were Trained And Evaluated On The ESP32-CAM: An Image-classification Model Distinguishing “Face” And “no_human” Classes, And An Object-detection Model Providing Spatial Localization. The Classification Model, Deployed With A 0.60 Confidence Threshold, Triggers Transmission Of The Captured JPEG Frame To A Telegram Bot For Remote Notification Whenever A Positive Detection Persists Beyond A Five-second Upload Cooldown. The Classification Model Achieved 96.0% Validation Accuracy, An F1 Score Of 0.97 For The Face Class And 0.95 For The No_human Class, And An Area Under The ROC Curve Of 0.95, While The Comparative Object-detection Model Achieved A Validation F1 Score Of 84.0% Over 60 Training Cycles. Experimental Results Confirm Functional Autonomous Navigation, Correctly Triggered Telegram Alerting And A Working End-to-end Pipeline. The Comparison Shows That Image Classification Is Best Suited To The Binary Alerting Requirement Of This Application, While Object Detection Offers Additional Spatial Information That Is Valuable For Future Localization-based Tasks.

Keywords