Case Study: Multi-Target Multi-Camera Tracking

Problem

Our client was a multi-national startup. They were aiming to build a robust yet cost-efficient system based on existing CCTV-style camera installation that could be used to track objects and people within the confines of an area.

Solution

TechnoLynx developed a solution that was both capable of identifying objects upon entering the area, as well as tracking them, both within the viewport of a single camera, as well as across cameras. As an additional challenge, during the project our team came to the realisation that the views of cameras do not always overlap with each other, yet having more cameras would impact the cost too much.

Our team overcame this obstacle by building a higher-level, probabilistic model of the path/trajectory of motion. With this additional predictive model, we reached an acceptable level of accuracy. The final solution was also compatible with a granular level of human supervision if needed.

Our team used a variety of software, including PyTorch, TensorRT, OpenCV, numpy and custom code in order to deliver the solution to cost-effective target hardware, both on edge for Raspberry PI (as an experiment) as well as an HPC-backed version.

Result

The resulting technology enabled the client to hold successful technology demonstrations to their leads in relevant operational environments, as well as to start collecting feedback directly from users. Our collaboration continued after these initial successes.