Customer wanted us to build an end-to-end model that enables marking and tracking a person or group of persons on a live video stream or multiple streams
PERSON TRACKING AND
RE-IDENTIFICATION
To build a robust recognition engines, following features need to be built :
- Face Detection : to identify the person who need to be tracked.
- Vector generator : for generating a unique vector for the person to be tracked.
FACE DETECTION FOR IDENTIFICATION
Home grown facial recognition system which is tweaked to provide the functionality of initial Detection and there of to identify / re-identify when a person moves out of frame and get back into the FOV
VECTOR GENERATION FOR TRACKING
Each person in the body will be tagged with a vector matrix.
TRACKING
Tracking algorithm works based on the input from recognition and tracks the person where in the frame irrespective of the person facing frontal or back to the camera.
Model seamlessly handles the situation where a person of interest going out of FOV and coming back into FOV and also, can track multiple people.
TECHNOLOGIES / TOOLS
- Keras, TensorFlow
- OpenCV
- Hybrid Model for Person Re-Identification and Tracking the person
MODEL HIGHLIGHTS
- Tested on the videos with few people to crowds on the streets
- Tested on both low light and good lighting conditions
- Can be used to train and track objects other than humans