kqtenqk / mot-deepsort-cs

Multiple object tracking

This is the C# implementation of SoRT and DeepSoRT trackers using YOLO network as person predictor and OSNet as appearance extractor in the frame.

Examples

SoRT example

DeepSoRT example

Console app usage (windows)

Command line options

  -s, --source        Required. Source video file path.

  -t, --target        Required. Target video file path.

  -d, --detector      Required. Detector net file path.

  -m, --matcher       Required. Matcher type:
                       0 for DeepSort
                       1 for Sort
                       2 for Deep

  -y, --yver          Required. Yolo model:
                       0 for 1280 resolution
                       1 for 640 resolution

  -v, --aver          Appearance model:
                       0 for OSNet
                       1 Fast-Reid

  -a, --appearance    Appearance extractor net file path.

  --fps               Target video fps.

  --threshold         Defines treshold for matcher.

  --aweight           Defines appearance weight for deepsort matcher.

  --asmooth           Defines appearance smooth weight for deepsort matcher.

  --streak            Defines min streak to reidentify person.

  --misses            Defines max misses to lose indentification.

  --fsmooth           Defines passed frames for smooth weight to be applied.

  --acount            Defines appearance extractors in memory count.

  -c, --conf          Defines target people detection confidence([0-1]).

  --help              Display this help screen.

  --version           Display version information.

Example of using

mot_x64 -s source.mp4 -t target.mp4 -d yolo640v5.onnx -y 1 -m 0 -a osnet.onnx -v 0 -c .4

MOT.CORE usage

Used file hierarchy:

. 
└ Assets
    ├ Input
    │   └ test.mp4
    ├ Output
    └ Models
        ├ Yolo
        │   └ yolo640v5.onnx
        └ Reid
            └ osnet_x1_0_msmt17.onnx

Some .onnx models are in src/MOT/ directory.

Initializing predictor and extractor

string yoloPath = "../../../Assets/Models/Yolo/yolo640v5.onnx";
var predictor = new YoloScorer<Yolo640v5>(File.ReadAllBytes(yoloPath));

string osnetPath = "../../../Assets/Models/Reid/osnet_x1_0_msmt17.onnx";
int extractorsInMemoryCount = 3;
var appearanceExtractor = new ReidScorer<OSNet_x1_0>(File.ReadAllBytes(osnetPath),
    extractorsInMemoryCount);

Initializing SoRT matcher

var matcher = new SortMatcher(predictor);

Initializing DeepSoRT matcher

var matcher = new DeepSortMatcher(predictor, appearanceExtractor);

Drawing people bounding boxed in the frame.

private static void DrawTracks(Bitmap frame, IReadOnlyList<ITrack> tracks)
{
    Graphics graphics = Graphics.FromImage(frame);

    foreach (ITrack track in tracks)
    {
        const int penSize = 4;
        const float yBoundingBoxIntent = 45f;
        const float xNumberIntent = 4f;
        const int fontSize = 44;

        graphics.DrawRectangles(new Pen(track.Color, penSize),
            new[] { track.CurrentBoundingBox });

        graphics.FillRectangle(new SolidBrush(track.Color),
            new RectangleF(track.CurrentBoundingBox.X - (penSize / 2), 
                track.CurrentBoundingBox.Y - yBoundingBoxIntent,
                track.CurrentBoundingBox.Width + penSize, 
                yBoundingBoxIntent - (penSize / 2)));

        (float x, float y) = (track.CurrentBoundingBox.X - xNumberIntent, 
                            track.CurrentBoundingBox.Y - yBoundingBoxIntent);

        graphics.DrawString($"{track.Id}",
            new Font("Consolas", fontSize, GraphicsUnit.Pixel), 
            new SolidBrush(Color.FromArgb((0xFF << 24) | 0xDDDDDD)),
            new PointF(x, y));
    }

    graphics.Dispose();
}

Getting video frame using Emgu.CV, handling frame and drawing bounding boxes

VideoCapture videoCapture = new VideoCapture("../../../Assets/Input/test.mp4");

double targetFps = videoCapture.Get(Emgu.CV.CvEnum.CapProp.Fps);
int width = videoCapture.Width;
int height = videoCapture.Height;

VideoWriter videoWriter = new VideoWriter("../../../Assets/Output/test.mp4", -1, 
                                    targetFps, new Size(width, height), true);

string yoloPath = "../../../Assets/Models/Yolo/yolo640v5.onnx";
var predictor = new YoloScorer<Yolo640v5>(File.ReadAllBytes(yoloPath));

string osnetPath = "../../../Assets/Models/Reid/osnet_x1_0_msmt17.onnx"; 
int extractorsInMemoryCount = 3;
var appearanceExtractor = new ReidScorer<OSNet_x1_0>(File.ReadAllBytes(osnetPath),
    extractorsInMemoryCount);

var matcher = new DeepSortMatcher(predictor, appearanceExtractor);
float targetConfidence = 0.4f;

Mat readBuffer = new Mat();

videoCapture.Read(readBuffer);

while (readBuffer.IsEmpty == false)
{
    Bitmap frame = readBuffer.ToBitmap();

    IReadOnlyList<ITrack> tracks = matcher.Run(frame, targetConfidence, DetectionObjectType.Person);

    DrawTracks(frame, tracks);

    videoWriter.Write(frame.ToImage<Emgu.CV.Structure.Bgr, byte>());
    videoCapture.Read(readBuffer);
}

matcher.Dispose();
videoWriter.Dispose();