Multi-Object Visual Tracking for Indoor Images of Retail Consumers

Visual tracking of multiple objects is a challenging task that involves several steps: target detection, data association and re-identification, each with its own set of challenges. Trending deep learning methods usually combine some of these steps and report impressive results in tracking benchmarks while running at practical speeds. In this work, we develop an online, real-time multi-object tracking approach which relies on a multi-task learning framework to output candidate targets and their corresponding discriminative image representations. The proposed model is also able to perform accurate retail consumer tracking without further tuning its design philosophy. Our key contribution lies at the motion forecasting part wherein standard Kalman Filters are replaced with a LSTM network which exploits long-term dependencies to essentially model target tracking. Numerical experiments of our method on the widely used MOT16 benchmark demonstrate its effectiveness. Additionally, quantitative results of in-house retail image data confirm the method's ability for practical multi-consumer tracking.