Introduction
This is a C++ implementation of SORT, a simple online and realtime tracking algorithm for 2D multiple object tracking in video sequences. Original Python code and publication infomation found at https://github.com/abewley/sort , By Alex Bewley
This code has been tested on Windows with Visual Studio Community 2013 + OpenCV 2.4.8. It depends on OpenCV without any other libs, so theoretically it can be compiled on Linux with OpenCV support.
Detection data in The ./data folder come from the original directory of SORT. They are the Faster RCNN detections for the MOT benchmark sequences in the benchmark format, created by Alex Bewley.
The Hungarian algorithm implementation comes from https://github.com/mcximing/hungarian-algorithm-cpp, which is derived from Markus Buehren's code.
The license is also kept to the GPLv3 License.
Results
The overall results of code are close to those of original SORT implmentation.
Using the MOT challenge devkit this code produces the following results.
| Sequence | Rcll Prcn FAR | GT MT PT ML | FP FN IDs FM | MOTA MOTP MOTAL |
|---|---|---|---|---|
| TUD-Campus | 67.4 91.0 0.34 | 8 5 3 0 | 24 117 8 12 | 58.5 74.8 60.5 |
| ETH-Sunnyday | 77.8 81.7 0.92 | 30 12 15 3 | 324 412 22 57 | 59.2 74.6 60.3 |
| ETH-Pedcross2 | 51.5 89.8 0.44 | 133 16 61 56 | 366 3037 83 118 | 44.3 74.6 45.6 |
| ADL-Rundle-8 | 44.3 75.3 1.50 | 28 6 15 7 | 984 3779 109 214 | 28.2 70.9 29.8 |
| Venice-2 | 42.6 64.7 2.77 | 26 8 11 7 | 1661 4098 63 112 | 18.5 72.8 19.3 |
| KITTI-17 | 67.9 92.6 0.26 | 9 1 8 0 | 37 219 13 16 | 60.6 72.6 62.4 |
| Overall | 49.5 77.1 1.28 | 234 48 113 73 | 3396 11662 298 529 | 33.5 73.1 34.8 |
A bit worse than original SORT but not significantly. This may due to my keeping the default parameters of Kalman filters. The speed of this implementation is at about 1800 FPS, or ~3 seconds over 5500 frames on my machine, with the Debug configuration.