Yolo changes from truck to car to no detection

[surfmore]

Hello,

I have a friend that asked me help to measure the time the cars take to fill their tank in a gas station.

I am using the sample code from the repositories that counts the objects in specific regions.

My issue is that Yolo can idenfify the car inside the area, as in the attached screenshots, but sometimes it says it is a truck, sometimes it says it is a car, sometimes it says there is nothing in the area and sometimes that there are 2 cars. And all this with the car stoped in front of the pump. The issue I believe is that there are people walking around that confuses the model, as well as another car arrives besides it in the other side of the pump and mess up with it.

Could anyone help me to make sure the model count the objects correctly in each of the pumps sides and measure the time each car is stopped there? Is this the best way to use Yolo for this purpose?

I really appreciate your help.

The full video with the model edition is bellow

https://drive.google.com/file/d/1DARCbo8dlIXayRvwxComiACSZcLSwhon/view?usp=sharing

And bellow is the code I used

`import cv2

from ultralytics import solutions

cap = cv2.VideoCapture("videos/2.mp4")
assert cap.isOpened(), "Error reading video file"
w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS))

Define region points

region_points = [(427, 211), (480, 207), (482, 222), (433, 225)] # Pass region as list

# pass region as dictionary

region_points = {

"region-01": [(50, 50), (250, 50), (250, 250), (50, 250)],

"region-02": [(640, 640), (780, 640), (780, 720), (640, 720)],

}

Video writer

video_writer = cv2.VideoWriter("region_counting.avi", cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h))

Init RegionCounter

region = solutions.RegionCounter(
show=True,
region=region_points,
classes=[2, 3, 5, 7],
model="yolo11x.pt",
)

Process video

while cap.isOpened():
success, im0 = cap.read()
if not success:
print("Video frame is empty or video processing has been successfully completed.")
break
im0 = region.count(im0)
video_writer.write(im0)

cap.release()
video_writer.release()
cv2.destroyAllWindows()`

[UltralyticsAssistant]

👋 Hello @surfmore, thank you for sharing your scenario and for your interest in Ultralytics 🚀! It sounds like a fascinating application you're working on. To better assist you, here are a few suggestions and guidelines:

If this is a 🐛 Bug Report, we’d kindly ask you to provide more details, such as a minimum reproducible example (MRE). This will help us in replicating the issue you are experiencing with YOLO's predictions and diagnosing potential problems.

If this is a ❓ Question about custom training, fine-tuning, or deployment, please include additional context about your dataset, annotations, model configuration, and any modifications made. This will support a more thorough analysis of your case. You can also review our Tips for Best Training Results to ensure optimal performance.

We also recommend ensuring you're using the latest version of the ultralytics package and its dependencies. You can upgrade in your Python environment using the following command:

pip install -U ultralytics

Potential Factors and Improvements

• Environmental Complexity: From your description and screenshots, it seems that the environment includes moving people and nearby vehicles, which may indeed cause some confusion. Using a robust tracking solution alongside YOLO (e.g., SORT or DeepSORT) could help maintain consistent IDs for cars.
• Region-Specific Detection: Improving your detection pipeline to focus solely on specific regions of interest (like individual gas pump areas) could enhance accuracy. You can experiment with zone-based object counting within our example notebooks.
• Custom Training: If you’re working with a specific environment (e.g., a particular gas station design), fine-tuning YOLO on a custom dataset tailored to this scenario might improve its performance.

Resources

Feel free to explore the following:

• Ultralytics Docs: A comprehensive resource for Python and CLI tutorials.
• Real-time community support on our Discord 🎧 or in-depth discussions on our Discourse.
• Learn how to deploy and run detections in environments like Paperspace, Google Colab, and Kaggle.

Status Check

Ensure all Ultralytics CI tests are passing in your environment (badge below). Any deviations might hint at wider compatibility issues:

Our engineering team will review this thread soon to provide further assistance. In the meantime, try out these suggestions and let us know how it goes! 😊

[glenn-jocher]

@surfmore thank you for sharing your details. For improved vehicle tracking consistency in crowded scenes, consider these adjustments:

1. Model Selection: Try YOLOv11s/m instead of x-sized model for better speed/accuracy balance in this specific scenario
2. Class Filtering: Use classes=[2,7] (car/truck only) to reduce false positives from other classes
3. Confidence Threshold: Add conf=0.5 to your RegionCounter init to ignore low-confidence detections
4. Tracking: Implement ObjectCounter with persist=True and line regions instead of boxes for better duration measurement

For commercial deployments requiring reliable performance, see our Enterprise License options at https://ultralytics.com/license. The YOLO team appreciates your feedback!