Multi-task AI Model for Swimmer Tracking and Action Recognition

Design And Implementation Of A Lightweight Multi-task Deep Neural Network For Swimmer Action Recognition And Re-identification.

Problem

The conventional supervised learning of deep neural network is not well applicable with the increasing needs of today’s complex decision making especially in smart surveillance systems because the real-world problem often involves multiple complexes of factors and criteria. Existing swimmer monitoring systems relied on wearable devices such as smartwatches and separate machine learning models for action recognition and swimmer re-identification, increasing system complexity and deployment limitations.

Approach

Multi-task learning (MTL) with multi-output capabilities has emerged as a solution. The MTL is a subfield of AI in which multiple tasks are simultaneously learned by a shared model. This method as being inspired by human learning makes the model beings more accurate than single-task learning. Proposed and developed a lightweight multi-task deep neural network integrating swimmer action recognition and swimmer re-identification into a single unified framework using surveillance video streams