Reidentifier
Reidentifier microservice processes feature vectors or clusters and finds reappearances.
If a processed feature vector or cluster is unknown for the system, it becomes registered. If the processed feature vector or cluster is already registered, it is reidentified by the system.
Camera input streams and already clustered streams can be used for registration, reidentification or both. The microservice can handle multiple input topics and produces results into one output topic. The roles of input topics, reidentification parameters, registration parameters, and topic consuming parameters can be configured.
The feature vectors or clusters belonging to a single individual are similar even
when changing the distance, head pose, and so on. For each pair of feature
vectors or clusters, a similarity score is produced. This score falls between
0.0 and 1.0. A higher score means, it is more likely that the two feature
vectors or clusters represent the same individual. Reidentifier is
recognizing this correlation between feature vectors and clusters based on
similarity scores.
The microservice has basically two functions:
Registers the new feature vectors or clusters coming from registration streams. Feature vectors and clusters coming from reidentification-only streams are never stored.
Tries to reidentify feature vectors and clusters coming from reidentification streams. The input is compared to all stored clusters and the similarity scores are computed; the result is controlled by configuration parameters.
Person Stream
The person stream is used to add feature vectors of previously known individuals to the internal database of the Reidentifier. Keys of the records in the person stream can be changed to any string, such as the name of the individual or a unique ID.
In general, the identity of the detected persons detected by the reidentifier is unknown when the source is simple camera stream. The person stream offers a solution to this by enabling the reidentifier to match the detected persons to a previously defined identity.
Person streams can also be regarded as feature vector databases, in case they are compacted. In this case, only one feature vector belongs to each key.
Further reading: