Yolov5 DeepSort Pytorch Versions Save

• Per class tracking now enabled for create_tracker
• Tag and release naming fixed

All trackers inherit BaseTracker

• Added scripts for ultra fast experimentation by pre-generating detections and embeddings such that they can later be fed to any tracking method.
• Added possibility to pass embeddings to all trackers.
• ~3x CI pipeline speedup by using parallelization and pregenerated detections and embeddings

with_reid argument was added to BoTSORT by @Kevin111369. Usage:

tracker = BoTSORT(
    model_weights = None,
    device = 'cuda:0',
    fp16 = True,
    with_reid = True,  # true for motion + ReID association, false for only motion association
) 

A centroid-based association method has been added to OCSORT and DeepOCSORT. This may be best suited for small AND/OR fast moving objects

from boxmot import OCSORT

tracker = OCSORT(
    asso_func="centroid",
    iou_threshold=0.3  # use this to set the centroid threshold that match your use-case best
)

A centroid-based association method has been added. This may be best suited for small AND/OR fast moving objects

HybridSort numpy datatypes fix by @florian-fischer-swarm

The KF adapter contained some minor bugs. All orginal KF from their respective repositories are now used instead.

BoTSORT improvements by @Justin900429 in https://github.com/mikel-brostrom/yolo_tracking/pull/1192:

• Integration of a Fusing Option
• Optimization of Optical Flow Computation

Increase reid mutlibackend preprocessing robustness by clipping the detections to ((0, w), (0, h)) by @Justin900429 in https://github.com/mikel-brostrom/yolo_tracking/pull/1187. This error may arise if the detector generates bboxes that ends outside the input image. It happens when the detector has been trained, not following best practices (clean up your ground truth, clip the output of the model...).