2023 - Full Presentation Schedule
Exploring Quantized CNN Deployment on Microcontrollers for Object Detection in a PAN
Start Date
31-3-2023 1:45 PM
End Date
31-3-2023 2:00 PM
Location
CASB 103 - Physical, Computer, and Chemical Science
Document Type
Presentation
Abstract
TinyML is a growing field that focuses on developing deep learning models suitable for ultra-low-power devices such as microcontrollers (MCU).
Traditional computer vision models use computationally expensive convolutional neural networks (CNNs), making even the quantized forms of these models difficult to deploy on microcontrollers. We present our efforts in designing a multi-node TinyML system for the detection of multiple objects. To accomplish this objective, we experimented with using camera modules with the ESPCAM and the NANO BLE 33 embedded devices.
We further explored using embedded devices' systems with one MCU per node and systems with multiple MCUs per node. The goal of our effort is to develop a system that performs object detection and notifies the user about the presence of an object within a personal area network (PAN).
When compared to a traditional network-based object-detection system, our proposed system provides the following advantages. First, our system is simple and cost-effective to deploy as it is completely based on a PAN and does not need Wi-Fi or Internet connectivity. Secondly, our system processes data inside the PAN, hence ensuring the privacy of surveillance data.
Exploring Quantized CNN Deployment on Microcontrollers for Object Detection in a PAN
CASB 103 - Physical, Computer, and Chemical Science
TinyML is a growing field that focuses on developing deep learning models suitable for ultra-low-power devices such as microcontrollers (MCU).
Traditional computer vision models use computationally expensive convolutional neural networks (CNNs), making even the quantized forms of these models difficult to deploy on microcontrollers. We present our efforts in designing a multi-node TinyML system for the detection of multiple objects. To accomplish this objective, we experimented with using camera modules with the ESPCAM and the NANO BLE 33 embedded devices.
We further explored using embedded devices' systems with one MCU per node and systems with multiple MCUs per node. The goal of our effort is to develop a system that performs object detection and notifies the user about the presence of an object within a personal area network (PAN).
When compared to a traditional network-based object-detection system, our proposed system provides the following advantages. First, our system is simple and cost-effective to deploy as it is completely based on a PAN and does not need Wi-Fi or Internet connectivity. Secondly, our system processes data inside the PAN, hence ensuring the privacy of surveillance data.