Real-time detection
Processes object detection continuously from a live webcam feed.
Personal computer vision project
TowTruck Alert
Learning Python, OpenCV, and modern object detection through a real-time vehicle detection and tracking pipeline.
- Status
- V1 Complete
- Focus
- Computer vision fundamentals
- Input
- Live webcam feed
Overview
The original idea was to detect tow trucks entering an apartment complex and eventually notify residents. As I worked through the problem, the project became a broader exercise in building a solid real-time object detection and tracking pipeline.
The finished application processes a webcam feed, detects selected vehicle classes with Ultralytics YOLO, draws confidence-scored boxes, calculates center points, and assigns persistent IDs as objects move through the frame.
Current capabilities
Object tracking
Uses YOLO tracking to maintain persistent IDs across frames.
Vehicle filtering
Displays selected classes such as cars, trucks, and buses.
Confidence scores
Shows the model confidence associated with each detection.
Bounding boxes
Uses OpenCV to highlight detected objects in the frame.
Center-point calculation
Calculates and displays object center coordinates.
Workflow
Webcam
↓ Frame capture
↓ YOLO detection
↓ Vehicle filtering
↓ Object tracking
↓ Bounding boxes
↓ Tracking IDs + confidence
↓ Display frame
Learning progression
01
First successful face detection
The first milestone: loading OpenCV, detecting a face with a Haar Cascade, and drawing a first detection box.
02
Finding my first major bug
A misunderstanding of video frames opened hundreds of image windows—a major debugging lesson.
03
Understanding video streams
Debugging made it clear that OpenCV continuously processes webcam frames instead of creating individual images.
04
Persistent object tracking
The final milestone: Ultralytics YOLO detections receive persistent tracking IDs while moving through the scene.
Screenshots
Challenges
OpenCV coordinate systems
Learning how image coordinates map to boxes, center points, and frame annotations.
Video-frame processing
Understanding that a live feed is a continuous stream rather than a sequence of static image windows.
YOLO output
Learning to interpret pretrained-model detections, classes, and confidence values.
Detection versus tracking
Moving from identifying an object in one frame to following it through a scene with a persistent ID.
What I learned
- Working with OpenCV and image coordinates.
- Drawing graphics and labels directly on video frames.
- Debugging real-time applications through small experiments.
- Using pretrained machine-learning models responsibly.
- Organizing a larger Python project into manageable milestones.
- The value of iteration when a first idea evolves into a stronger learning project.
Technical stack
Takeaway
Although TowTruck Alert began with a specific tow-truck use case, it evolved into a much broader exploration of computer vision. The project gave me a practical foundation for future AI and vision work: how a frame becomes a detection, and how a detection becomes something you can follow over time.
Explore the repository ↗