DeepDish on a diet

Abstract

Intelligent sensors using deep learning to comprehend video streams have become commonly used to track and analyse the movement of people and vehicles in public spaces. The models and hardware become more powerful at regular and frequent intervals. However, this computational marvel has come at the expense of heavy energy usage. If intelligent sensors are to become ubiquitous, such as being installed at every junction and frequently along every street in a city, then their power draw will become non-trivial, posing a severe downside to their usage. We explore Multi-Object Tracking (MOT) solutions based on our custom system that use less power while still maintaining reasonable accuracy.