Joachim Hummel
504f8c0b69
- Linienschnitt-Algorithmus für präzise Fahrzeugzählung - Interaktive Linienauswahl im Browser (Canvas-basiert) - Session-Management für benutzerdefinierte Zähllinien - Typ-spezifische Zähler (Autos, LKW, Busse, Motorräder) - REST-API für Linienkonfiguration und Zähler-Reset - Gestrichelte Zähllinie als Video-Overlay - Detailliertes Zähler-Display im Video Features: - Linienüberquerung-Erkennung (beide Richtungen) - Keine Mehrfachzählung durch Track-ID-Management - Funktioniert für Webcam und Video-Upload - Benutzerfreundliche UI mit Echtzeit-Feedback 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
87 lines
3.5 KiB
Markdown
87 lines
3.5 KiB
Markdown
# CLAUDE.md
|
|
|
|
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
|
|
|
|
## Project Overview
|
|
|
|
This is a Flask-based web application for real-time object detection and tracking using YOLOv11. The application supports both webcam streaming and uploaded video file processing with object detection overlays.
|
|
|
|
## Key Dependencies
|
|
|
|
- **Flask**: Web framework for routing and serving HTML templates
|
|
- **OpenCV (cv2)**: Video processing and frame manipulation
|
|
- **Ultralytics YOLO**: YOLOv11 model for object detection and tracking
|
|
- **NumPy**: Array operations for image data
|
|
|
|
The YOLO model file (`yolo11s.pt`) must be present in the root directory.
|
|
|
|
## Running the Application
|
|
|
|
Start the Flask development server:
|
|
```bash
|
|
python3 app.py
|
|
```
|
|
|
|
The application runs on `0.0.0.0:8080` and is accessible at `http://localhost:8080`
|
|
|
|
## Application Architecture
|
|
|
|
### Single-File Structure
|
|
The entire application is contained in `app.py` with no separate modules. All routes, video processing logic, and model initialization are in this single file.
|
|
|
|
### Core Components
|
|
|
|
**YOLO Model (lines 10-11)**
|
|
- Model loaded once at application startup: `model = YOLO("yolo11s.pt")`
|
|
- Class names extracted from model for labeling: `names = model.model.names`
|
|
|
|
**Video Processing Pattern**
|
|
Both webcam and uploaded video processing share identical detection logic:
|
|
1. Frame skipping: Only processes every 2nd frame (`count % 2 != 0`) for performance
|
|
2. Frame resizing: All frames resized to (1020, 600) for consistent output
|
|
3. YOLO tracking: Uses `model.track(frame, persist=True)` to maintain object IDs across frames
|
|
4. Annotation: Draws bounding boxes and labels with format `{track_id} - {class_name}`
|
|
5. Streaming: Yields JPEG frames as multipart HTTP response
|
|
|
|
**Duplicate Code**
|
|
The functions `detect_objects_from_webcam()` (lines 21-52) and `detect_objects_from_video()` (lines 86-118) contain nearly identical logic - only the video source differs (webcam vs. file path).
|
|
|
|
### Routes and Flow
|
|
|
|
**Main Routes:**
|
|
- `/` - Landing page (index.html) with video upload form and webcam button
|
|
- `/start_webcam` - Webcam detection page (webcam.html)
|
|
- `/upload` [POST] - Handles video file uploads, saves to `uploads/` directory
|
|
- `/uploads/<filename>` - Video playback page (play_video.html)
|
|
|
|
**Streaming Routes:**
|
|
- `/webcam_feed` - Streams processed webcam frames with detections
|
|
- `/video_feed/<filename>` - Streams processed video file frames with detections
|
|
- `/video/<filename>` - Serves raw video files from uploads directory
|
|
|
|
### Templates
|
|
All templates follow the same pattern:
|
|
- Centered layout with flex containers
|
|
- Fixed dimensions (1020x600) for video display
|
|
- "Back to Home" link for navigation
|
|
|
|
## Important Notes
|
|
|
|
**File Uploads**
|
|
- Uploaded videos are saved to `uploads/` directory (created automatically if missing)
|
|
- No file size limits or validation beyond checking file existence
|
|
- No cleanup mechanism for uploaded files
|
|
|
|
**Performance Optimization**
|
|
- Frame skipping (every 2nd frame) reduces CPU load but may miss fast-moving objects
|
|
- Frame resizing to 1020x600 is hardcoded throughout
|
|
|
|
**Object Tracking**
|
|
- Uses persistent tracking (`persist=True`) to maintain object IDs across frames
|
|
- Track IDs are displayed alongside class names in the format: `{track_id} - {class_name}`
|
|
- Tracking state persists within a single video stream session
|
|
|
|
**Webcam Access**
|
|
- Uses `cv2.VideoCapture(0)` for default webcam
|
|
- No error handling if webcam is unavailable or already in use
|