Vehicle Detection Computer Vision

I built Vehicle_detection as an educational and portfolio project for detecting and counting vehicles in images and videos. My Jupyter Notebook project uses Python and OpenCV for image and video processing, and the Roboflow Workflow API for vehicle predictions. The README names me as Aakash Kumar, B.Tech CSE – Data Science, and describes the project as started and last updated in 2026-09. It is also listed among my computer-vision projects on github profile.

• I use the system to detect car, truck, bus, bicycle, and motorcycle classes. It can process images and videos, detect multiple vehicles, visualize bounding boxes, display classes and confidence scores, and count detections. The project also uses Base64 image encoding, REST API requests, and frame-by-frame video processing.

For image detection, I read an image with OpenCV, encode it as Base64, and send it in a JSON request to the Roboflow Workflow API. I extract the returned JSON predictions, ignore predictions below the 0.30 confidence threshold, convert each prediction’s center coordinates and width and height into box corners, and draw green OpenCV rectangles with thickness 3. I display the class and confidence and save the result as vehicle_detection_result.jpg.

For video, I read the video properties and process frames with OpenCV, sending every 5th frame to Roboflow to reduce API requests. I draw the latest predictions, write processed frames, and generate vehicle_detection_output.mp4, which includes boxes, labels, confidence scores, and vehicle counts. Because the project does not describe tracking, I do not treat these counts as confirmed counts of unique vehicles; skipping frames may also miss short-lived detections.

payload = {
    "api_key": API_KEY,
    "inputs": {
        "image": {
            "type": "base64",
            "value": image_base64
        },
        "classes": "car, truck, bus, bicycle, motorcycle"
    }
}

response = requests.post(
    API_URL,
    headers={"Content-Type": "application/json"},
    json=payload
)

I keep the Roboflow credential in ROBOFLOW_API_KEY in a local .env file, load it with dotenv, and raise ValueError("ROBOFLOW_API_KEY not found") if it is missing. I include .env in .gitignore and do not push it to GitHub. If a key has already been exposed publicly, I revoke it and generate a replacement before making the repository public.

I run Vehicle_detection.ipynb in Google Colab, Jupyter Notebook, or VS Code with the Jupyter extension. The documented repository includes the notebook, README.md, requirements.txt, .gitignore, .env, and an output/ directory. The README clone command uses YOUR_USERNAME, so it does not establish the complete repository URL. It also does not specify the workflow endpoint, workflow identifier, model details, API plan, dataset, or evaluation results.

The project depends on the Roboflow API and an internet connection; available API-plan limits and the time needed to send video frames to an external service can constrain processing. Detection performance also depends on image quality, lighting, camera angle, and the underlying model. I list vehicle tracking, license-plate detection, speed estimation, traffic-density analysis, entry/exit counting, line-crossing detection, a local YOLO model, video-processing optimization, and online deployment as possible future improvements.

Building this project gave me practice with computer-vision and object-detection fundamentals, OpenCV image and video processing, REST and JSON API integration, Base64 encoding, bounding-box coordinate conversion, confidence thresholding, video-frame processing, and vehicle counting. The README acknowledges the Roboflow Workflow API for detection and OpenCV and Python for image and video processing.