OPTIMIZATION OF YOLOV8 FAMILY MODELS TO WORK ON DEVICES WITH LIMITED COMPUTING RESOURCES WHEN DETECTING THE CONDITION OF PARKING SPACES
Authors:
1.
Kazan National Research Technical University named after A.N. Tupolev
(Department for Automated Systems for Information Processing and Control, Assistant)
employee from 01.01.2024 to 01.01.2025
employee from 01.01.2024 to 01.01.2025
2.
Kazanskiy nacional'nyy issledovatel'skiy tehnicheskiy universitet imeni A. N. Tupoleva-KAI
Russian Federation
Russian Federation
Type:
Article
Pages:
from 89
to 92
Status:
Published
Received:
25.12.2025
Accepted:
25.12.2025
Published:
25.12.2025
Language:
Russian
Keywords:
COMPUTER VISION, DEEP LEARNING, YOLO, QUANTIZATION, OBJECT DETECTION, OPTIMIZATION OF NEURAL NETWORKS
Abstract (English):
Today, the development of so-called "Smart Parking" systems is associated with the rational use of parking space. Automatic parking monitoring systems are aimed at solving the urgent problem of efficient allocation of car owners' time. A parking space is a collection of individual spaces - areas intended for parking vehicles. Usually such areas are marked up in advance, however, this article presents a special case of a parking space without appropriate marking lines, so the task of the monitoring system in this case is to detect parking spaces and recognize their states (occupied or vacant). The tasks of object detection and recognition can be combined into an object detection task. The article discusses the optimization of neural networks of the YOLOv8 family to effectively perform the task of detecting the state of parking spaces on devices with limited computing resources.. The main focus is on the use of static and dynamic quantization methods to reduce the size of the model and speed up its operation while maintaining acceptable accuracy. A comparative analysis of the methods of these approaches is presented. Experiments demonstrate that quantization of weights and activations can significantly reduce the computational complexity and memory required for YOLOv8 family models to be deployed on devices with limited computing resources. Despite the theoretical advantages of the adaptive approach, the dynamically quantized model showed worse results in terms of processing speed (1.2 FPS) compared to the static method. Also, after static quantization, the model provides an average processing speed of 2.0 FPS, (1.7 times faster than the dynamically quantized version (1.2 FPS) and 2.2 times faster than the original model (0.9 FPS)). At the same time, the total video processing time for the statically quantized model was significantly less than in other variants (665.25 seconds).
Today, the development of so-called "Smart Parking" systems is associated with the rational use of parking space. Automatic parking monitoring systems are aimed at solving the urgent problem of efficient allocation of car owners' time. A parking space is a collection of individual spaces - areas intended for parking vehicles. Usually such areas are marked up in advance, however, this article presents a special case of a parking space without appropriate marking lines, so the task of the monitoring system in this case is to detect parking spaces and recognize their states (occupied or vacant). The tasks of object detection and recognition can be combined into an object detection task. The article discusses the optimization of neural networks of the YOLOv8 family to effectively perform the task of detecting the state of parking spaces on devices with limited computing resources.. The main focus is on the use of static and dynamic quantization methods to reduce the size of the model and speed up its operation while maintaining acceptable accuracy. A comparative analysis of the methods of these approaches is presented. Experiments demonstrate that quantization of weights and activations can significantly reduce the computational complexity and memory required for YOLOv8 family models to be deployed on devices with limited computing resources. Despite the theoretical advantages of the adaptive approach, the dynamically quantized model showed worse results in terms of processing speed (1.2 FPS) compared to the static method. Also, after static quantization, the model provides an average processing speed of 2.0 FPS, (1.7 times faster than the dynamically quantized version (1.2 FPS) and 2.2 times faster than the original model (0.9 FPS)). At the same time, the total video processing time for the statically quantized model was significantly less than in other variants (665.25 seconds).
Keywords:
COMPUTER VISION, DEEP LEARNING, YOLO, QUANTIZATION, OBJECT DETECTION, OPTIMIZATION OF NEURAL NETWORKS
COMPUTER VISION, DEEP LEARNING, YOLO, QUANTIZATION, OBJECT DETECTION, OPTIMIZATION OF NEURAL NETWORKS
