Project Description
I have a fixed-angle camera that watches every truck as it rolls through a single gate into our yard. What I need is a reliable, end-to-end Python pipeline that will:
• Detect and track each individual truck in real time with a pre-trained YOLO model, keeping the ID stable from the moment the vehicle enters the frame until it leaves.
• Within that per-truck track, run additional YOLO passes (or custom classes) to locate the key regions I care about: license plates, DOT numbers, chassis numbers, container numbers, and container type markings. Accuracy on these regions is critical; false positives must be minimal.
• Crop each detected region, perform OCR, and then clean the raw text with solid post-processing logic—regex, checksums, or any heuristic that standardises spelling, spacing, and character sets. The final output for each truck should clearly list:
– License Plate
– DOT Number
– Chassis Number
– Container Number
– Container Type
• Automatically create a new folder named with the truck’s unique track ID and timestamp. Inside that folder save:
– The original frame and cropped regions
– A JSON (or CSV) file containing the cleaned text fields
– Any confidence scores you deem useful for auditing
Key expectations
1. The solution must run offline on a Linux box with Python, OpenCV, and any OCR engine you feel is best (Tesseract is fine, but I’m open to suggestions).
2. Processing speed has to keep up with 30 fps footage from a 1080p stream; batching or frame-skipping strategies are acceptable if they don’t hurt reliability.
3. Code quality matters: well-structured modules, clear function names, and a concise README explaining environment setup, command-line arguments, and sample output.
4. Provide a short video demo or screen-capture proving end-to-end detection, OCR, and folder creation.
Deliverables
• Fully-commented source code and requirements.txt
• Pre-trained weights or a link to them plus any fine-tuning notebook you created
• README with setup steps, run commands, and troubleshooting notes
• A recorded demo confirming the pipeline meets the above criteria
If you have experience combining YOLO tracking with OCR and text clean-up, this should be a straightforward but interesting challenge. Let me know how you plan to tackle false positives and maintain ID consistency, and we’ll get started right away.