Generating labeled data for deep object tracking

A system comprising:

• a processor;
• and
• a memory coupled to the processor, the memory storing computer-executing instructions that, when executed by the processor, cause the system to:
• for an object of interest identified in a given temporal frame included in a sequence of temporal frames:
• calculate a likelihood function to correlate a feature score of the object of interest with respective feature scores associated with each one of one or more target identifiers (IDs) previously assigned to objects of interest in a selected range of the sequence of temporal frames, wherein each temporal frame in the selected range of the sequence of temporal frames is a previously annotated temporal frame;
• assign a target ID to the object of interest by:
• identifying a previously assigned target ID associated with a highest value of the calculated likelihood function and assigning the identified target ID to the object of interest;
• or
• initiating a new target ID when none of the previously assigned target IDs is associated with a value of the calculated likelihood function that satisfies a predetermined threshold and assigning the initiated new target ID to the object of interest;
• and
• perform track management according to a predefined track management scheme to assign a track type to the object of interest;
• iterate the calculating, assigning and performing for all objects of interest identified in the given temporal frame;
• annotate the given temporal frame to include all target IDs and track types assigned to all objects of interest identified in the given temporal frame;
• and
• add the annotated given temporal frame to an annotated dataset, wherein the annotated dataset includes previously annotated temporal frames, each previously annotated temporal frame including target IDs and track types assigned to all objects of interest in the previously annotated temporal frame.

A non-transitory computer-readable medium comprising instructions which, when executed by a processing unit cause the processing unit to generate an annotated temporal frame for an annotated dataset used for training a deep tracking neural network by:

• for an object of interest identified in a given temporal frame included in a sequence of temporal frames:
• calculating a likelihood function] to correlate a feature score of the object of interest with respective feature scores associated with each one of one or more target identifiers (IDs) previously assigned to respective objects of interest in a selected range of the sequence of temporal frames, wherein each temporal frame in the selected range of the sequence of temporal frames is a previously annotated temporal frame;
• assigning a target ID to the object of interest by:
• identifying a previously assigned target ID associated with a highest value of the calculated likelihood function and assigning the identified target ID to the object of interest;
• or
• initiating a new target ID when none of the previously assigned target IDs is associated with a value of the calculated likelihood function that satisfies a predetermined threshold and assigning the initiated new target ID to the object of interest;
• and
• performing track management according to a predefined track management scheme to assign a track type to the object of interest;
• iterating the calculating, assigning and performing for all objects of interest identified in the given temporal frame;
• annotating the given temporal frame to include all target IDs and track types assigned to all objects of interest identified in the given temporal frame;
• and
• adding the annotated given temporal frame to the annotated dataset, wherein the annotated dataset includes previously annotated temporal frames, each previously annotated temporal frame including target IDs and track types assigned to all objects of interest in the previously annotated temporal frame.