yasenh / sort-cpp Goto Github PK

Hello @yasenh , it's a nice work, however i am bit optimistic about adding functionality to show trace lines for the tracked object, you can just look into this kalman_filter_multi_object_tracking SORT implementation which is in python and does pretty good job in tacking object along with the trace lines.

Reproduced on g++ 7.5.0, g++ 10.3.0, clang 11.0.0.

Sample leads to tracker.Run() infinity cycle:

#include <opencv2/core.hpp>

#include "tracker.h"

int main(int argc, const char *argv[]) {

    Tracker tracker;

    std::vector<std::vector<cv::Rect>> all_detections = {
        {cv::Rect(15, 4, 12, 2)},
        {cv::Rect(15, 4, 6, 2)},
        {cv::Rect(15, 4, 2, 2)},
        {cv::Rect(12, 5, 5, 2)},
    };

    for (auto det : all_detections) {
        tracker.Run(det); // On the last `det` munkers algorithm stucks in infinity cycle.
    }

}

Reason:

Kalman filter prediction hasn't zero bounds on width and height (in this code block problem in width, see logs below). So track object that is built from Kalman prediction can be with negative width/height.

Track area will be negative if one of the track's width/height is negative and other is positive. So it can be case (like in code block) when
formula in Tracker::CalculateIou float union_area = det_area + trk_area - intersection_area; equals zero and auto iou = intersection_area / union_area; got zero devision which leads iou==(+/-)nan.

Some logs:

1)
Track::ConvertStateToBbox rect:[12 x 2 from (15, 4)]
Tracker::CalculateIou det_area:12 trk_area:24 intersection_area:12 iou:0.5

2)
Track::ConvertStateToBbox rect:[0 x 2 from (15, 4)]
Tracker::CalculateIou det_area:4 trk_area:0 intersection_area:0 iou:0

3)
Track::ConvertStateToBbox rect:[-5 x 2 from (14, 4)]
Tracker::CalculateIou det_area:10 trk_area:-10 intersection_area:0 iou:-nan  << !!!!!

4)
Track::ConvertStateToBbox rect:[2 x 2 from (15, 4)]
Tracker::CalculateIou det_area:10 trk_area:4 intersection_area:2 iou:0.166667

Infinity cycle here because Munker's matrics have `-nan` from step 3.

Pull request with bugfix will be soon.

'M_PI': undeclared identifier in kalman_filter.cpp(116,83)

Hi
Thanks for your great work. I've written python bindings and a package for SORT algorithm using this repository. Other python packages which I had found were implemented in python and performed slowly, so I decided to use C++ behind and wrote this module. It also can be installed via pip.

https://github.com/MrGolden1/sort-python

If you wish, include the link in your README.

Thank you!

@yasenh can this source code track objects having orientation ?? if not where can i add the orientation of the object parameters ??

@yasenh thanks for wonderful source code , i had few quereis

1. the performance results are low compared to https://github.com/mcximing/sort-cpp
2. is this the similar implementation or different
3. what all fine tuning has to be done in the code
4. will this work 3d object tracking also

Thanks in advance

deepsort
?

It seems