IoT enabled object detection on edge devices
Tasmi, Tahsina Ferdous (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025052817433
https://urn.fi/URN:NBN:fi:amk-2025052817433
Tiivistelmä
Manual quality assurance for checking each product and process in large-scale production systems is error-prone, time-consuming, and costly. Therefore, the focus of this thesis was to research and develop an automatic quality assurance and product maintenance system utilizing IoT measuring systems, data collection, and machine learning models on resource-limited embedded edge units to detect different objects on a conveyor belt in the manufacturing processes. Two different types of batteries were utilized to demonstrate object detection of different products on the same assembly line.
The implementation integrated the custom TensorFlow Lite (int8 quantized) FOMO (Faster Object More Object) machine learning model deployed on the edge computing device (microcontroller unit) with the OV2640 camera sensor. The aim was to automatically detect different types of batteries on the same assembly line based on the trained datasets which can help to ensure proper maintenance and quality. The custom machine learning model deployed on the resource-constrained edge device was able to specifically detect the exact type of each product in real-time streaming. Each component of the project served an important part to enable the system to collect data, process, and monitor products on the assembly line.
Even though some technical challenges were encountered during the developmental phase, the final system produced high-performance, useful, and mostly accurate results with a 96% F1 score in detecting different objects according to their specific types. The system demonstrated the feasibility of detecting multiple objects in order to ensure proper maintenance. It managed to handle multiple challenges such as energy efficiency, high processing speed, and resource constraints. The performance of handling multiple object detection tasks automatically with minimal manual intervention highlights the potential of numerous useful applications.
FOMO (Faster Object More Object) is one of the fastest and highest-performing machine learning models that could be implemented on resource-constrained edge units such as microcontrollers. FOMO was successfully utilized in this research project in order to automate multi-object detection tasks in the context of real-time product monitoring in manufacturing processes.
This research project demonstrated a significant and advanced integration of lightweight machine learning models on the microcontroller unit in order to facilitate intelligence and computation nearer to the data, enabling real-time processing with low latency, and improved privacy. It offered several benefits such as high performance and increased efficiency. Additionally, it offered advanced technical integration of IoT and machine learning models on edge units as part of the Industry 4.0 principles.
The implementation integrated the custom TensorFlow Lite (int8 quantized) FOMO (Faster Object More Object) machine learning model deployed on the edge computing device (microcontroller unit) with the OV2640 camera sensor. The aim was to automatically detect different types of batteries on the same assembly line based on the trained datasets which can help to ensure proper maintenance and quality. The custom machine learning model deployed on the resource-constrained edge device was able to specifically detect the exact type of each product in real-time streaming. Each component of the project served an important part to enable the system to collect data, process, and monitor products on the assembly line.
Even though some technical challenges were encountered during the developmental phase, the final system produced high-performance, useful, and mostly accurate results with a 96% F1 score in detecting different objects according to their specific types. The system demonstrated the feasibility of detecting multiple objects in order to ensure proper maintenance. It managed to handle multiple challenges such as energy efficiency, high processing speed, and resource constraints. The performance of handling multiple object detection tasks automatically with minimal manual intervention highlights the potential of numerous useful applications.
FOMO (Faster Object More Object) is one of the fastest and highest-performing machine learning models that could be implemented on resource-constrained edge units such as microcontrollers. FOMO was successfully utilized in this research project in order to automate multi-object detection tasks in the context of real-time product monitoring in manufacturing processes.
This research project demonstrated a significant and advanced integration of lightweight machine learning models on the microcontroller unit in order to facilitate intelligence and computation nearer to the data, enabling real-time processing with low latency, and improved privacy. It offered several benefits such as high performance and increased efficiency. Additionally, it offered advanced technical integration of IoT and machine learning models on edge units as part of the Industry 4.0 principles.