Object Detection from Cloud to Edge
Järvinen, Kasper (2022)
Järvinen, Kasper
2022
Tietojohtamisen DI-ohjelma - Master's Programme in Information and Knowledge Management
Johtamisen ja talouden tiedekunta - Faculty of Management and Business
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2022-11-28
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202211248627
https://urn.fi/URN:NBN:fi:tuni-202211248627
Tiivistelmä
This thesis studies the differences of object detection between cloud computing and edge computing systems. The objective is to create a view on how fast and accurately object detection can be done using novel edge computing technologies, compared to cloud computing approaches. The research uses literature review to create a view of the current state of cloud computing and edge computing paradigms and real-time object detection. Then, an empirical test is performed to test the performance of a selected edge hardware using real-world data. The perks of the test system are further discussed together with the findings of the empirical test.
First in the literature review, cloud computing and edge computing paradigms are studied. Their basic components and logic are presented, together with the main challenges emerging from them. Then, object detection as an application of machine learning and computer vision is reviewed. The research then briefly represents the functioning logic of two distinct real-time object detection models, that are later used in the empirical test of this work. Finally, ways to measure object detection performance are explored.
In the empirical test, two selected object detection models are implemented into the selected edge hardware. Testing data is then applied to the machine by emulating an IP camera stream with a record of traffic video. Metrics and measurements, such as inference times and CPU utilization are then sent to a cloud server, giving insight into the performance of the device using each model. The architecture includes an edge hardware, that has an integrated cloud service, through which the application and device management is done.
In the light of the empirical study’s results, it is found that new edge computing hardware can do object detection in real-time, still resulting in slower inference times than cloud solutions that have practically unlimited computing resources. However, when considering for example the data transfer latencies when using cloud instances, the overall object detection process performed at the edge can compete with the cloud solutions.
The empirical test set up followed a new approach to the edge paradigm, using an integrated service between the cloud and edge. When comparing this process further to the obstacles found in cloud and edge computing technologies, some improvements and partly solutions can be identified. These include improvements in security, latency, and delivery processes.
First in the literature review, cloud computing and edge computing paradigms are studied. Their basic components and logic are presented, together with the main challenges emerging from them. Then, object detection as an application of machine learning and computer vision is reviewed. The research then briefly represents the functioning logic of two distinct real-time object detection models, that are later used in the empirical test of this work. Finally, ways to measure object detection performance are explored.
In the empirical test, two selected object detection models are implemented into the selected edge hardware. Testing data is then applied to the machine by emulating an IP camera stream with a record of traffic video. Metrics and measurements, such as inference times and CPU utilization are then sent to a cloud server, giving insight into the performance of the device using each model. The architecture includes an edge hardware, that has an integrated cloud service, through which the application and device management is done.
In the light of the empirical study’s results, it is found that new edge computing hardware can do object detection in real-time, still resulting in slower inference times than cloud solutions that have practically unlimited computing resources. However, when considering for example the data transfer latencies when using cloud instances, the overall object detection process performed at the edge can compete with the cloud solutions.
The empirical test set up followed a new approach to the edge paradigm, using an integrated service between the cloud and edge. When comparing this process further to the obstacles found in cloud and edge computing technologies, some improvements and partly solutions can be identified. These include improvements in security, latency, and delivery processes.