drone-detection-yolov5

🚁 Drone Detection & Real-Time Tracking — YOLOv5

"Can a $35 computer see, think, and chase a drone — all at once?"
That was the question. This is the answer.

---

🎯 The Challenge

Drones are everywhere. Detecting them in real-time on powerful hardware is one thing — but what if you only have a Raspberry Pi 4? No GPU, no CUDA, no shortcuts.

The challenge was clear:

• Detect drones accurately in real-time
• Run inference at usable frame rates on a CPU
• Move a physical camera to follow the drone — automatically

Most people would say it's not possible. We built it anyway.

---

📊 Final Model Performance

---

💡 The Story Behind This Model

Act 1 — The First Attempt

We started simple. A single dataset, ~500 images, YOLOv5s with Transfer Learning. The first model came out at 95% mAP — impressive on paper.

But on the Raspberry Pi, reality hit hard.

The camera would lock onto a wooden tray and call it a drone. A chair. A lamp. The model was too eager — it had never truly learned what wasn't a drone.

We had a precision problem. And we knew exactly why: not enough negative images.

---

Act 2 — The Dataset Problem

The original dataset had only 5 background images out of 503 total. Five. The model had barely seen what the world looks like without a drone in it.

So we went hunting for data.

Instead of spending weeks manually collecting and labeling images, we combined 3 of the largest drone datasets on Roboflow Universe:

But combining datasets isn't just copy-paste. Each dataset has its own format, resolution, and labeling style. We built a custom pipeline to:

1. Download all 3 datasets via Roboflow API
2. Convert every label to unified YOLO format
3. Auto-label our positive images using the first model as a labeling tool
4. Add COCO background images as hard negatives

It took engineering. It took iteration. But it worked.

---

Act 3 — Training at Scale

With 40,000+ images ready, we needed serious compute. We upgraded to an NVIDIA A100 40GB GPU on Google Colab and trained for 30 epochs with:

Model     : YOLOv5s (Transfer Learning from COCO)
Image size: 416×416
Batch size: 256
Epochs    : 30
GPU       : NVIDIA A100 40GB
Time      : ~45 minutes

The results were immediate. mAP jumped from 95% to 99%. False positives dropped from 100% to 1%. The model had finally learned the difference between a drone and everything else.

---

Act 4 — Running on the Edge

A great model means nothing if it can't run where it needs to. We exported to ONNX format and deployed on Raspberry Pi 4 using ONNXRuntime — no GPU, no PyTorch, just pure optimized inference.

Achieved: 15-20 FPS on a $35 computer. ✅

---

Act 5 — The Tracker

Detection alone wasn't enough. We built a complete pan-tilt tracking system:

Camera Feed → YOLOv5 Detection → PID Controller → Servo Motors

A 4-state machine manages the behavior:

SEARCHING → drone appears → TRACKING
TRACKING  → drone lost   → LOST
LOST      → timeout      → RETURNING
RETURNING → centered      → SEARCHING

The PID controller computes servo movement based on the pixel error between the drone's center and the frame center:

Pan  axis: Kp=0.055  Ki=0.0005  Kd=0.010
Tilt axis: Kp=0.045  Ki=0.0005  Kd=0.010
Deadzone : 10 pixels (prevents micro-jitter)

Hardware stack:

• Raspberry Pi 4 — main compute
• USB Webcam — 640×480 live feed
• PCA9685 PWM Driver — I2C servo controller
• 2× Servo Motors — pan and tilt axes
• External 5V/2A PSU — dedicated servo power

---

🔧 Quick Start

Install

pip install onnxruntime opencv-python numpy

Detect on Image

import cv2
import numpy as np
import onnxruntime as ort

session    = ort.InferenceSession("best.onnx")
input_name = session.get_inputs()[0].name

def letterbox(frame, size=416):
    h0, w0 = frame.shape[:2]
    r      = min(size/h0, size/w0)
    nw, nh = int(round(w0*r)), int(round(h0*r))
    dw, dh = (size-nw)/2, (size-nh)/2
    img    = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
    img    = cv2.resize(img, (nw, nh))
    img    = cv2.copyMakeBorder(img,
                int(round(dh-.1)), int(round(dh+.1)),
                int(round(dw-.1)), int(round(dw+.1)),
                cv2.BORDER_CONSTANT, value=(114,114,114))
    img = np.expand_dims(img.transpose(2,0,1), 0).astype(np.float32)/255.0
    return img, r, dw, dh

def detect(image_path, conf_thres=0.45):
    frame      = cv2.imread(image_path)
    h0, w0     = frame.shape[:2]
    img, r, dw, dh = letterbox(frame)
    preds      = session.run(None, {input_name: img})[0][0]
    mask       = preds[:, 4] > conf_thres
    preds      = preds[mask]
    results    = []
    for det in preds:
        cx   = int((det[0] - dw) / r)
        cy   = int((det[1] - dh) / r)
        w    = int(det[2] / r)
        h    = int(det[3] / r)
        conf = float(det[4])
        x1   = max(0,  cx - w//2)
        y1   = max(0,  cy - h//2)
        x2   = min(w0, cx + w//2)
        y2   = min(h0, cy + h//2)
        results.append({'bbox': [x1,y1,x2,y2], 'conf': conf})
        cv2.rectangle(frame, (x1,y1), (x2,y2), (0,255,0), 2)
        cv2.putText(frame, f"Drone {conf:.0%}", (x1, y1-10),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255,255,255), 2)
    cv2.imwrite("output.jpg", frame)
    print(f"Detected {len(results)} drone(s)")
    return results

results = detect("your_image.jpg")

Detect on Live Camera

cap = cv2.VideoCapture(0)
while True:
    ret, frame = cap.read()
    if not ret: break
    img, r, dw, dh = letterbox(frame)
    preds = session.run(None, {input_name: img})[0][0]
    mask  = preds[:, 4] > 0.45
    for det in preds[mask]:
        cx = int((det[0]-dw)/r); cy = int((det[1]-dh)/r)
        w  = int(det[2]/r);      h  = int(det[3]/r)
        cv2.rectangle(frame, (cx-w//2, cy-h//2), (cx+w//2, cy+h//2), (0,255,0), 2)
    cv2.imshow("Drone Detector", frame)
    if cv2.waitKey(1) & 0xFF == ord('q'): break
cap.release()

---

📁 Repository Files

---

👤 About

Author: Ahmed Darwish
Deployment: Raspberry Pi 4
Training: Google Colab — NVIDIA A100 40GB GPU
Stack: Python · YOLOv5 · ONNXRuntime · OpenCV · PCA9685 · PID Control

---

Built from scratch. Trained at scale. Deployed on the edge. 🚀