Person Re-Identification in Different Cameras

• State-of-the-art accuracy in person re-identification
• Real-time
• Robust to changes in person gait, appearance, pose, illumination and camera orientation
• The trained neural network only minorly reduces its re-identification accuracy when applied on the multi-camera setups different from the training multi-camera setup

Input(s):

• Timestamped video streams from multiple cameras
• Bounding boxes of detected pedestrians
• UUID of each detected pedestrians

Main feature(s):

• The component is able to detect a person and find his/her previous appearances in the recordings from other cameras in a multi-camera network

Output(s):

• For each queried pedestrian: all previous detections in different cameras, i.e. UUID of each previous detection
• Timestamped moving trajectory of a pedestrian

• Accurate timestamping of the captured video is required
• SW constraints: no
• HW constraints: NVIDIA GPU, 8 GB GPU RAM

ViNotion, Police, surveillance and security operators, Research community (via open source).

Licensing