Ελληνικό Μεσογειακό Πανεπιστήμιο
Έξυπνες συσκευές στο διαδίκτυο των πραγμάτων με επεξεργασία στα άκρα του δικτύου
Abstract
dc:descriptionΙnsects play a vital role in ecosystems, participating in pollination and other natural processes, however they can cause problems for agricultural crops. Therefore, it is necessary to control and regulate their populations. Traditional techniques for monitoring insects have a limited range, are time-consuming, and require many skilled workers. Advances in technology, along with the development of the Internet of Things and the rapid evolution of Artificial Intelligence, make it possible to monitor insects automatically and non-invasively with smart traps. This thesis focuses on the use of Internet of Things and edge computing technologies to develop smart devices that incorporate artificial intelligence techniques. The goal is to create autonomous, innovative real-time data monitoring systems that meet the modern needs of agriculture and environmental management. Among the key contributions of the thesis is the development of a low-cost electronic insect trap (smart e-trap) that performs on-site image processing with a built-in deep learning model for automatic counting of harmful insects. A cost-effective device was developed, which has increased energy autonomy and provides satisfactory image quality for identification purposes. This device is designed to count the number of insects and transfer images from inside the trap to a remote server. The analysis and counting of insects is performed locally, using built-in quantized deep learning models. This study compares various performance parameters of three different edge devices, namely ESP32, Raspberry Pi Model 4, and Google Coral, which run deep learning algorithms using the TensorFlow Lite development framework. Each edge device was able to process images and achieve an accuracy of over 95%. However, processing times and energy consumption varied between the different devices. Our findings suggest that ESP32 emerges as the most suitable choice in the context of this application in line with our policy for low-cost and low-power devices.The second part describes the development of an image-based insect trap with features that distinguish it from other camera traps: (a) no manual labeling of training images is required to learn how to count target insects, and (b) it autonomously removes captured insects, making it ideal for prolonged use. The device includes a funnel trap with pheromone, into which a Raspberry Pi microcontroller equipped with a camera and deep learning algorithms for counting agricultural pests is integrated. In addition, the device receives instructions from the server that determines its functionality, and a built-in servo motor can automatically rotate the removable bottom of the bucket to dispose of dehydrated insects when they begin to accumulate. As a result, it overcomes a significant limitation of camera-based insect traps, which is the inevitable clogging and obstruction caused by the accumulation of insects over time, enhancing the trap’s autonomous functionality. Cases that are not adequately covered in the literature are examined, such as counting in crowded conditionsand large concentrations using crowd counting algorithms, especially in human surveillance scenarios. Finally, a comparative analysis of the results obtained from insect population measurements is carried out, applying different methods with deep learning algorithms: a) object detection algorithms (YOLOv7/8), b) algorithms for crowd counting (CSRNET), and c) regression algorithms in convolutional models. It is noteworthy that there is no single, ideal measurement method suitable for all circumstances involving overlapping insects of different sizes. After evaluating the advantages and disadvantages, it is suggested that YOLOv7/8 is generally the most satisfactory solution. In the third part, the new Diopsis dataset, which includes annotated images from sticky traps in the field, is used to examine an approach that focuses on enhancing the image set withdifferent backgrounds, as well as the SAHI approach that allows the detection of microscopic objects in object detection models such as YOLO. Extracting insects from training images to create enhanced backgrounds and using these backgrounds as bases for artificially shifting insects can enhance detection accuracy. In addition, SAHI techniques are combined during the operation of the YOLOv10nano model, achieving an mAP50 of 72.7%. Achieving mAP50 at 72.7% using a small model such as YOLOv10nano, which is suitable for small edge devices, demonstrates the efficient tracking of insects on energy-efficient devices such as the Raspberry Pi Zero W 2. Finally, extending the scope of this work beyond biodiversity monitoring practices and visual counting, we propose the development of a comprehensive acoustic monitoring system that combines a variety of technologies. This system aims to create a platform for distributed monitoring of audio data in real time. The system uses a network of microphones mounted on ESP32 devices, which transmit compressed audio clips via MQTT protocol to Raspberry Pi 5 devices for acoustic classification. These devices incorporate an audio transformer (AST) model trained on the AudioSet dataset, which allows for immediate categorization and time-stamping of audio events with high accuracy. The output of the transformer is stored in an event database and then converted to JSON format upon request or interface requirements. The generated JSON files are further processed into a graph structure, which incorporates the annotated soundscape, offering a multidimensional and holistic representation of sound environments. The graphs and JSON structures are then examined and analyzed using special Python code and language models (LLM), allowing the system to respond to complex questions regarding the nature, relationships, and context of the recorded sound events.A new graph analysis method is proposed that achieves low false alarm rates. The methodology for creating and processing event graphs outlined above was applied to the analysis of the audio content of a 1 hour and 40 minute film featuring dangerous driving practices. Our graph-based approach achieved an accuracy of 0.882, precision of 0.8, recall of 1.0, and an F1 score of 0.89.By combining the reliability of distributed sound detection with the accuracy of transformer-based sound classification (AST), the proposed method treats sound data as text, laying the foundations for new, advanced applications in acoustic surveillance and environmental monitoring, among others.
Degree
thesis:*- Grantor dc:publisher
- Ελληνικό Μεσογειακό Πανεπιστήμιο
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sarantopoulos, Ioannis
- Σαραντόπουλος, Ιωάννης
Subjects
dc:subject × 31- Ηλεκτρονικές παγίδες
- Ανίχνευση παρασίτων
- Καταμέτρηση εντόμων
- Αισθητήρες εικόνας
- Ακουστική επιτήρηση
- Τεχνητή νοημοσύνη
- Βαθιά μάθηση
- Τεχνητά νευρωνικά δίκτυα
- Υπολογιστική άκρης
- Συσκευές άκρης
- Τεχνιτή νοημοσύνη άκρης
- Διαδίκτυο των πραγμάτων
- Εlectronic traps
- Pest detection
- Insect counting
- Image sensors
- Acoustic surveillance
- Audio surveillance
- Artificial Intelligence (AI)
- Deep learning
- Artificial neural networks
- Edge computing
- Internet of things
- Edge devices
- Edge AI
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Αυτοματισμοί και Συστήματα ελέγχου
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Automation and control systems
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60172