Tracking with frame skips

[abewley]

Tracking with frame skips

[abewley]

To clarify the comment about calling the update() method even with empty detections. This is because the tracker makes a constant velocity assumption and also will delete object if too many frames have passed without seeing them. By default this is set to 1 so if you don't supply detections at least every second frame the objects will be lost. See lines 226-227.

In the case when you are sub-sampling frames and maintaining a fixed frequency you could simply treat the image sequence as a different framerate and you do not need to call it between frames when you are not doing any detection. Although your code appears to already be doing this.

[TheMikeyR]

For other people who encounter same issue as me:

After some debugging it got to my attention that the min_hits=3 will remove the tracks if i feed empty detections in,since the min hit only increase when the function is fed detection, so 3 frames with detections in a row will make the counter go to 3, and the updated tracks will be returned. https://github.com/abewley/sort/blob/master/sort.py#L222

Since I'm trying to feed it every frame, and only feed detections every x frame it requires min_hits=1, so it is able to give tracks back. To solve it either edit the sort.py default value at https://github.com/abewley/sort/blob/master/sort.py#L176 or when creating the object from your script you can give it a new value as input like tracker = Sort(min_hits=1)

Hope this information will help people who encountered the same issue and couldn't figure out a solution.

[kkyiss]

Hey @TheMikeyR
I want to try track with some detect frame skip ,and I have tried your idea above which suggest setting
tracker = Sort(min_hits=1)
but still once I feed empty list, the tracker output returns empty

Then I read over the @abewley commented, which said

By default this is set to 1 so if you don't supply detections at least every second frame the objects will be lost. See lines 226-227.

also with #17 Smmothing detections using tracking , the commented below said

I changed the following line to https://github.com/abewley/sort/blob/master/sort.py#L222
if((trk.time_since_update < self.max_age+1) and (trk.hit_streak >= self.min_hits or self.frame_count <= self.min_hits)):

and set min_hits=0 and max_age=4

In my opinion, the line "By default this is set to 1" is refer to parameter max_age
and if we increment the value, the tracker can tolerate every q frames with detection information feed

At the end, by adding self.max_age to Line222 and set
tracker = Sort(max_age=5, min_hits=0)

I could track with every q frames skip (e.g q = 1 or 2 or 3 or 4 )
if q is larger than 5, the max_age has to be larger

not sure I'm doing right, do I miss anything compare to your idea
thanks for any replies

[TheMikeyR]

@kkyiss really good point, And I'm sorry I've totally missed to add info about max_age parameter, I believe it is due to the reason that I modified it early in my process and forgot that I modified mine when I tried to get frame skip to work. Mine is set to max_age=10 which makes it able to restore tracks to the old id for a longer timer, since the old id is saved for 10 frames (it can also be dangerous to keep a long history/age, but it seems to work fine with my data).

It looks like you are doing the same steps as I did, so I believe you are doing it right! Thanks for your comment, it will definitely help others who try to achieve the same thing!

[TheMikeyR]

Arhh, I see! Thanks a lot for your time.

[vivek-singh-rathore]

I am also trying to skip some frames for detection. But I am having problem in sending empty detections to sort's update function. It gives error in this line when I send an empty list of detections :

trk.update(dets[d,:][0])

IndexError: too many indices for array

[rishabh700]

I understand the role of max_age in the code. It helps to keep the tracker_id alive for the set number of frames and hence would reduce the number of new ids assigned to the same object.

But I am not sure what max_hits is doing? Does it set the minimum number of consecutive detections hits needed to assign the detection a new id if not matched with any other id? Please correct me if I am wrong!