🚗 Traffic Analysis & Vehicle Speed Estimation

YOLOv8 + ByteTrack + Perspective Transform (Bird’s-Eye-View)

This project performs vehicle detection, tracking, type labeling, and real-world speed estimation from a monocular traffic camera. It uses:

• YOLOv8 for object detection
• ByteTrack for multi-object tracking
• Supervision for annotation + utilities
• OpenCV Perspective Transform to convert camera view → Bird’s-Eye-View (BEV)
• Homography-based speed estimation for accurate km/h values
• Polygon ROI filtering for lane-specific measurement

🖼️ Visual Outputs

• Bounding boxes with vehicle types
• Trace lines showing motion history
• Speed labels (km/h)
• Valid lane polygon overlay
• Exported annotated video

---

📌 Key Features

• ✔️ Vehicle detection (car, truck, bus, bike…)
• ✔️ Multi-object tracking with persistent IDs
• ✔️ Real-world speed estimation (km/h)
• ✔️ Perspective → Bird’s-Eye transformation
• ✔️ Road-lane polygon filtering
• ✔️ Class-label display (#3 car 78 km/h)
• ✔️ Bounding boxes, traces, annotations
• ✔️ Export annotated output video
• ✔️ Ready for multi-video batch processing

---

🎥 How Speed Estimation Works

The core idea comes from mapping the camera’s distorted trapezoidal road area to a rectangular BEV space, where every pixel equals a fixed real-world distance.

1. Define SOURCE points (camera view)

Four points manually selected around the drivable area. These form a trapezoid because of perspective.
(Source diagram on page 2 of the PDF oai_citation:1‡NOTEBOOK PDF - Traffic - Computer Vision.pdf)

2. Define TARGET rectangle (flat BEV space)

A clean rectangle representing real-world dimensions.

These form a trapezoid because of perspective.
(Source diagram on page 2 of the PDF oai_citation:1‡NOTEBOOK PDF - Traffic - Computer Vision.pdf)

2. Define TARGET rectangle (flat BEV space)

A clean rectangle representing real-world dimensions:

3. Compute homography matrix

• This computes a 3×3 projective warp matrix mapping : SOURCE → TARGET.

4. Transform each tracked vehicle point

Now each vehicle’s movement occurs in uniform meter-like coordinates, enabling accurate km/h estimation.

5. Compute speed

distance = |y_start − y_end|
time = frames / FPS
speed(km/h) = (distance / time) * 3.6