Person Tracking with Intel's AI Reference Kit

This lab focuses on implementing live person tracking using Intel's OpenVINO™, a toolkit for high-performance deep learning inference. The objective is to read frames from a video sequence, detect people within the frames, assign unique identifiers to each person and track them as they move across frames. The tracking algorithm utilized here is Deep SORT (Simple Online and Realtime Tracking), an extension of SORT that incorporates appearance information along with motion for improved tracking accuracy.

Deep SORT, the tracking algorithm employed, combines motion and appearance information to effectively track objects. It comprises three main components: detection, prediction and data association/update. 

• In the detection phase, a deep learning model identifies objects within frames, passing these detections to the next stage.
• The prediction step employs the Kalman filter framework to anticipate the next bounding box for each tracked object, distinguishing between confirmed and unconfirmed tracks.
• Data association and update involve matching predicted bounding boxes with detections, updating track identities accordingly. This is achieved through a combination of metrics, including the Mahalanobis and cosine distances, for motion and appearance information integration, respectively. Intersection over Union (IOU) association is also applied to handle sudden appearance changes and increase robustness against errors, ensuring accurate tracking even in challenging scenarios.

• To understand the benefits of OpenVINO and Deep SORT
• Leverage Intel's Xeon 5th Generation processors to train, tune and inference the model

• Intel 5th Generation Processors on Dell PowerEdge R660 Servers
• OpenVINO
• Deep SORT
• OpenShift AI