Embedded AI and Human-Sensing Techniques in Intelligent Buildings
Metwaly, Aly (2019-09-12)
Embedded AI and Human-Sensing Techniques in Intelligent Buildings
Metwaly, Aly
(12.09.2019)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
suljettu
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2019100331037
https://urn.fi/URN:NBN:fi-fe2019100331037
Tiivistelmä
There is a rising demand for occupancy detection and people counting innovations for intelligent buildings. Occupancy detection capability is a promising reliable real-time solution for lighting control. People counting can be utilized in space management, smart workplaces, and market research. Furthermore, the HVAC systems can be adjusted according to the occupancy information which can lead to a dramatic reduction in the energy consumption of buildings. In Europe, buildings are considered the largest energy consumer using approximately 40% of the total energy and creating about 36% of the total CO2 emissions. Therefore, the energy reduction in this sector has a significant impact on the environment.
Previous research of people counting showed several limitations including low accuracy, high latency, and compromising the people's privacy. In this work, two sensing techniques were studied using several Machine Learning algorithms to test their suitability to the people counting application. These sensing techniques produce a plethora of data that needs to be collected, crunched and analyzed. Transferring this data to the center of the system increases the processing latency and reduces reliability because of its dependency on the network conditions. Moreover, it can have a significant impact on the network load if multiple sensors are utilized in the same building. Therefore, This research follows recent trends in more distributed computing and perform the AI computation on the embedded devices at the edge of the system. This approach is called Embedded AI.
In Embedded AI the analysis process is more reliable and provides lower latency. Moreover, data security and privacy are enhanced. However, this new approach comes with its challenges. The embedded systems have constrained resources such as low computational power, low memory, and low power consumption. The proposed novel solutions for people counting use different algorithms. The various algorithms offer flexibility as they have diverse MCU requirements. Therefore, the system can be tailored based on the available MCU resources.
The proposed method achieves an accuracy of 98.90% when predicting the number of people in an office room. This result significantly improves the current state-of-the-art accuracy. Furthermore, these algorithms have low computational and memory requirements. Therefore, the algorithms are suitable for resource-constrained devices such as IoT devices.
Previous research of people counting showed several limitations including low accuracy, high latency, and compromising the people's privacy. In this work, two sensing techniques were studied using several Machine Learning algorithms to test their suitability to the people counting application. These sensing techniques produce a plethora of data that needs to be collected, crunched and analyzed. Transferring this data to the center of the system increases the processing latency and reduces reliability because of its dependency on the network conditions. Moreover, it can have a significant impact on the network load if multiple sensors are utilized in the same building. Therefore, This research follows recent trends in more distributed computing and perform the AI computation on the embedded devices at the edge of the system. This approach is called Embedded AI.
In Embedded AI the analysis process is more reliable and provides lower latency. Moreover, data security and privacy are enhanced. However, this new approach comes with its challenges. The embedded systems have constrained resources such as low computational power, low memory, and low power consumption. The proposed novel solutions for people counting use different algorithms. The various algorithms offer flexibility as they have diverse MCU requirements. Therefore, the system can be tailored based on the available MCU resources.
The proposed method achieves an accuracy of 98.90% when predicting the number of people in an office room. This result significantly improves the current state-of-the-art accuracy. Furthermore, these algorithms have low computational and memory requirements. Therefore, the algorithms are suitable for resource-constrained devices such as IoT devices.