PRIVACY-AWARE EDGE COMPUTING SYSTEM FOR PEOPLE TRACKING

This work studies a practical implementation of a distributed computer vision system for people tracking. A particular focus is on improved data privacy when compared to the traditional surveillance approaches. This is achieved by extracting a feature vector from each detected person by a neural network in real-time in the edge device and transmitting only the feature vector to the cloud, eliminating privacy-sensitive image data transmission and storage. The proposed solution is implemented in a network of Raspberry Pi single board computers and Intel Neural Compute Stick accelerators. The system is tested in an environment where multiple edge devices are sending data to the cloud server for further analysis. In this context, we consider the spectrum of design and implementation aspects of real-time execution of multiple neural networks in a capacity limited edge computing environment.