Machine Learning Marker

The machine learning marker uses neural networks to detect the pedestrians. Therefore, the type of pedestrians that are detected depends on what the machine learning model was trained on. To keep it flexible, the user has to provide a model himself. For example, if he provides a model that was only trained on people wearing colored hats, then only pedestrians with colored hats are detected.

Downloads

If you don’t want to provide your own model, below there are 2 pretrained models compatible with PeTrack.

• Model trained on pedestrians without hats: download

• Model trained on pedestrians in experiments wearing different types of hats: download

Selecting the model version

Model version selection

After selecting machine learning as your preferred method for recognition, you have to choose which pretrained model your model is based on. Currently supported are YOLOv5 and YOLOv8-YOLOv11.

Machine learning parameter window

Machine learning parameter window

The parameter window lets you upload your custom model and tune various parameters to ensure optimal performance.

confidence threshold

The model provides a value for each detection describing how confident it is in its prediction. The confidence threshold determines how confident the model has to be in its prediction for it to be accepted as a valid detection.

NMS threshold

NMS (Non-Maximum Suppression) is a method to prevent multiple detections of the same object. It measures how much overlap there is between two detections. If it overlaps more than the specified threshold, then the detection with the lower confidence value is discarded.

score threshold

Besides a confidence value, the model also provides a value on how certain it is that a given detection belongs to a certain class. The score threshold determines how high this value has to be for a detection to be accepted. When using the default model there is only one class (Pedestrian), so this option has no effect. It is only useful when you differentiate between multiple classes of pedestrians.

image width

The image width refers to the image size used when training the model. When you use the default model, 640 is already the correct value.

model

Here you can upload your preferred model for recognition. The default models can be found in the Downloads section. The model has to be in the .onnx format and either be trained on YOLOv5 or YOLOv8.

class names

Here you can upload a .names file that lists all the used classes line by line, if you have more than one. When using the default model, you don’t have to upload anything, and Pedestrian is used as the only class.