Background

Over the past several years, there has been a large shift towards telemedicine, digital health and in-home fitness. The pandemic has only helped to accelerate this change in our approach away from in-person healthcare and exercise. Artificial intelligence plays a critical role in enabling technologies in these fields. One such technology is Pose Estimation, a computer vision task using a machine learning model to estimate the pose of a person from an image or video by estimating the special locations of key body joints. Pose estimation requires high computational resources which has limited the range of applications where it can be used. Traditionally, pose estimation is conducted via a graphical model. Recently, developments in deep Convolutional Neural Networks (CNN) have significantly boosted performance. Many researchers tend to use a deep and high-capacity CNN architecture pre-trained on large-scale datasets, and adapt it to the pose estimation task. However, the scaling problem still remains as a bottleneck. One popular solution is to seek a large capacity network as the backbone to learn better feature representations. Nevertheless, the model complexity of the large backbone architectures, in terms of time and space, makes it hard for direct applications on mobile and embedded devices. Therefore, there is a need to design customized light-weight CNN modules to reduce the computational cost and storage size for further on-device applications.

Technology Overview

Northeastern spinout, AInnovation, has invented a comprehensive artificial intelligence framework through novel light-weight deep learning techniques of quantitatively visual sensing, analyzing, understanding, and modulating human behaviors through low-cost and portable computer vision devices, such as smartphones and tablets. This framework enables complex computer vision programs, including human pose tracking, behavioral recognition, and motion capturing, to run on platforms constrained by limited hardware, bringing cutting-edge innovation to telemedicine, remote rehab, virtual-first care and in-home fitness.

Most existing light-weight model compression technology, has been focusing on the lower level or hardware level implementations, e.g. pruning, quantization or knowledge distillation. Such methods are generic and not application oriented, which may suffer considerable degrading of algorithm accuracy, and unpredictably fail in many cross-domain scenarios. In contrast, AInnovation has invented a customized AI model compression technology which is application oriented. The model design is implemented on the model components level, which is more robust and can achieve superior accuracy enhancement and model size reduction.

Benefits

Superior accuracy and robust pose estimation.
Computationally less expensive compared to the other pose estimation methods
Can run on resource constrained devices such as smart phones and tablets

Applications

Telemedicine & remote physiotherapy
In-home fitness
Gaming
Security/defense
E-learning/remote education
Any industry that requires an AI-based solution

Opportunity

Partnering for customized solutions
Identifying new space for company

Cutting-edge AI Innovation for On-Device Real-time Human Tracking and Pose Estimation

Patents

IP Status

Seeking