Due to the success of research on deep learning methods, the last ten years have seen an explosion in the development of artificial intelligence (AI) techniques for a wide variety of applications. Deep learning broadly refers to a class of algorithms that mimic the human brain structure and can be used to build systems that learn from previous data. Techniques based on deep learning have allowed ever more large and sophisticated machine learning models to be built and deployed, allowing a very rich set of complex problems to be solved. However, deep learning is computationally expensive, severely limiting its use on resource-constrained devices like single-board computers. Clarkson University Professors of Electrical and Computer Engineering Faraz Hussain and James Carroll along with Ph.D. student M. G. Sarwar Murshed have been working on designing novel techniques for deploying deep-learning-based intelligent systems in resource-constrained settings by optimizing models that can be used on edge devices.
Based on research supported by Badger Inc, they recently published a paper entitled “Resource-aware On-device Deep Learning for Supermarket Hazard Detection” in the 19th IEEE International Conference on Machine Learning and Applications (ICMLA ’20), which demonstrated a method for deploying deep learning models on small devices such as the Coral Dev board, Jetson Nano, and the Raspberry Pi. The paper describes a new dataset of images for supermarket floor hazards and a new deep learning model named EdgeLite to automatically identify such hazards, specifically intended to be used in extremely resource-constrained settings. EdgeLite processes all the images locally to allow it to be used to monitor supermarket floors in real-time. By processing all data locally using a resource-constrained device, EdgeLite helps preserves the privacy of the data.
A comparison of EdgeLite with six state-of-the-art deep learning models (viz. MobileNetV1, MobileNetV2, InceptionNet V1, InceptionNet V2, ResNet V1, and GoogleNet) for supermarket hazard detection when deployed on the Coral Dev Board, the Raspberry Pi, and the Nvidia Jetson TX2, showed it to have the highest accuracy and comparable resource requirements in terms of memory, inference time, and energy.
Further, they have successfully demonstrated how to deploy EdgeLite on autonomous robots. This was done using the Robot Operating System (ROS), a widely-used middleware platform for building autonomous robot applications. Using EdgeLite, a robot can identify hazardous floors by analyzing the image data without the help of additional hardware such as Lidar or other sensors, which can help a robot navigate through the supermarket aisles and report potential hazards, thus significantly improving safety.