search

Smorodov / Multitarget-tracker

Multiple Object Tracker, Based on Hungarian algorithm + Kalman filter.
Apache License 2.0
2.14k stars 646 forks source link

CarsCouting Problem #295

Closed 17520299 closed 2 years ago

17520299 commented 3 years ago

Excuse me, profressor. Cars Counting included measured velocity of vehicle function? I haven't installed yolo_lib yet.Instead, I have a trained-model car detection by Yolo v4. So i can use it for Cars Counting? Thanks you, sir!

Nuzhny007 commented 3 years ago

Hi!

  1. You can use google translate for our russian wiki: https://github.com/Smorodov/Multitarget-tracker/wiki/Fine-tuning-TrackerSettings

  2. I'm using different scores for different cases. For example for tracking a small objects like birds or UAVs the best metric is a DistCenters.

  3. Now, resolution reduce don't give you a big performance. For best detection a small vehicles You need set maxCropRatio = 1 (but it will slow). Also you can use TensorRT backend with batch size = 2 or 4 and FP16 precision on GPU 20xx or 30xx

  4. Yes, IoU is not enough. For this case used DeepSORT. In this project you can use it here with re-id models from OpenVINO (your OpenCV must be compiled with OpenVINO support): https://github.com/Smorodov/Multitarget-tracker/blob/master/example/examples.h#L682

17520299 commented 3 years ago

Thank you very much for this response!!!

17520299 commented 3 years ago

Hi prof! Excuse me, can you explained me why period = min(2m_fps, traceSize) What does it mean? (3.6f m_fps)/period is time of object appear in real-time, isn't it? why 3.6f not another numbers? Thanks you so much, prof!!!

Nuzhny007 commented 3 years ago

Excuse me, can you explained me why period = min(2*m_fps, traceSize)

2 seconds for speed calculation. This number must be small for calculating instantaneous velocity and big for error minimization. I'm using 2 seconds.

(3.6f * m_fps)/period is time of object appear in real-time, isn't it? why 3.6f not another numbers?

3.6f - coefficient for converting velocity from meters/sec to km/h. 3.6 = 3600 seconds in hour / 1000 meters in km

17520299 commented 3 years ago

Hi, prof! when i use KalmanLinear and KalmanUnscented. With KalmanLinear vehicles are still tracked as below while KalmanUnscented not, and objects moved out frame but the detection still appear in frame. linear Linear unscented Unscented 00000159 Detection still appear for several frame while object moved out I read your Fine-tuning TrackSettings, you mentioned KalmanUnscented included acceleration more than Linear. Some articles say UnscentedKalman for non-linear, so i dont understand about what main different between KalmanLinear and KalmanUnscented in this case. Please explain me!!! Thank you so much, prof!

Nuzhny007 commented 3 years ago

I'm not sure that all my code with Uncented Kalman filer is correct. Linear is better

17520299 commented 2 years ago

Hi, prof!!! I have a question about DistRects. Which case you will use DistRects? and what is its reason? Please answer me when you see my question!!!! Please!!! Thanks you so much!!!

Nuzhny007 commented 2 years ago

I used this distance for tracking small objects with background subtraction based detector. It works better that centers dist and IoU (Jaccard) distance for very small objects doesn't work.

17520299 commented 2 years ago

Thanks you so much prof for support me up to now! So i will close this issue. Thanks you!!!