Internship @KLASS Engineering and Solutions

Throughout the 5-month internship at KLASS Engineering and Solutions, assignments related to video analytics were allocated. These assignments generally add new pipelines or enhancements to KLASS's video analytics software stack, KAvision.

KAvision is an analytics engine built around Deepstream by KLASS's engineers. KAvision includes multiple pipelines, each having its purposes, such as face recognition, mask detection, and much more.

KLASS's client had requested several video pipelines to use on their robotics platform. Some assignments allocated include developing two of the requested pipelines, namely the fashion classifier pipeline and the thermal person detection pipeline.

Assignments

Most of the assignments given during this internship aim to improve KLASS's video analytics software stack and create video analytics pipelines requested by the company's clients.

Assignments include adding a three-stage inference pipeline to detect clothing on varying body parts and creating a thermal person detector pipeline to detect humans in varying lighting conditions. These pipelines are a project deliverable requested by the company's client for their robotics platform. Other assignments include further improvements to the thermal person detector model and a tool to generate annotation files for entire datasets. This new tool would be an addition to KLASS's video analytics software stack and help engineers evaluate models, create new datasets and much more.

Output of video analytics pipeline consisting of body parts detector, clothing classifier and face recognition

Output of video analytics pipeline using thermal-person detector

By making use of the open-source darknet framework, new models were trained using different sets of configurations. After testing various architectures and utilizing various methods to improve the model, the final model trained using YOLOv4-Tiny architecture had significantly outperformed the original YOLOv3 model in terms of accuracy and performance.

The complete comparison between the various models trained for the thermal person detector

Internship Poster

Poster detailing one of the assignments completed this internship