Hi @fellipezambrano! Both issues you're experiencing are common tracking challenges. Setting confidence to 0.01 is actually causing more problems than solving them - this extremely low threshold creates noisy detections that confuse the tracker.

Try increasing your confidence to 0.3-0.5 and experiment with different tracking algorithms. You can specify the tracker with model.track(source=video, tracker="bytetrack.yaml") or tracker="botsort.yaml". ByteTrack often performs better in scenarios with multiple similar objects like cars, while BoT-SORT handles erratic movement patterns well.

For the double detection issue, you might also want to adjust the IoU threshold for NMS by adding iou=0.5 to your track parameters. If you're still getting inconsistent results, consider custom training on your specific camera angle and lighting conditions, as this can significantly improve tracking stability for your use case.

The hardware difference between YOLO11s and YOLO11m might not be worth it if your tracking issues aren't resolved first - focus on getting stable detections before optimizing for speed.

Hello @glenn-jocher, Thanks for your kinddle support.

While I waited some answer here in forum, I kept doing my research. Copilot has been very useful for me. He suggested I use this file YAML, as you with extra configuration, which last test (I did just one - my solution need to count cars in a drive trhu of fast food, so I need to wait the costumers).

With the recomendations of copilot, I adjusted my YAML, as below:

tracker_type: botsort
with_reid: True
model: auto
gmc_method: sparseOptFlow
fuse_score: True
match_thresh: 0.6

track_high_thresh: 0.40
track_low_thresh: 0.01
new_track_thresh: 0.5
track_buffer: 600
proximity_thresh: 0.3
appearance_thresh: 0.6

Share with me, please your thoghs about that. As I said, last night it performed well, but was just the first test.

About making a training with Yolo, as you suggested, my qyestion is: How many samples of cars should I take to be better than the standart model of a car?! Maybe I'm wrong but, as a car is a very common object to detect, I think that it would be very hard, to my model getting better than the Yolo standart. M'I correct?

Again, thanks.
Best regards.

Your YAML configuration looks much better! The track_high_thresh: 0.40 is a significant improvement over 0.01. However, I'd suggest raising track_low_thresh to around 0.15-0.20 and reducing track_buffer to 30-60 frames to prevent keeping dead tracks too long in a drive-thru scenario where cars move through quickly.

For custom training in your drive-thru use case, you're partially correct that cars are well-represented in standard datasets, but custom training can still be valuable for your specific conditions. You'd typically need 1000-2000 images minimum, but the real benefit comes from training on your exact camera angle, lighting conditions, and the specific types of vehicles that frequent your location. Since drive-thru cameras often have unique overhead angles and lighting challenges, custom training could improve consistency even with well-known objects like cars.

That said, try optimizing your tracking parameters first before investing time in custom training - your current approach with the improved YAML might solve your counting issues without the additional complexity.