| Summary: | The dissertation at hand provides a comprehensive investigation into the optimization of vehicle marker recognition efficiency, a critical element in various applications such as road detection, autonomous driving, and local unmanned combat. It specifically focuses on the exploration and improvement of the YOLOv4-tiny detection algorithm for efficient and accurate vehicle marker recognition.
The core of this dissertation revolves around the effective detection of targets in natural scenes, which is paramount for real-time decision-making systems. The ability of these systems to respond promptly and accurately is a function of the efficiency of target detection, and in this context, vehicle marker recognition.
The YOLO family of detection algorithms, particularly YOLOv4-tiny, is analyzed for their performance in various detection scenarios. The study reveals that different versions of the YOLO models yield variable results due to factors such as dataset variability, obstacle occlusion, and angle-based recognition inaccuracies.
To address these issues and enhance the vehicle marker recognition efficiency of the YOLOv4-tiny algorithm, two optimization techniques were implemented. The Mosaic data enhancement method was employed to improve recognition for obscured targets. Simultaneously, a multi-angle dataset was incorporated to counter the issue of low recognition accuracy at varying angles. These two methods improved the recognition accuracy without compromising the real-time response of the original model, hence enhancing the overall recognition efficiency.
Experimental validations were conducted using a camera-captured matchmaking scene dataset on a single-card platform NVIDIA GeForce MX450. The findings demonstrated a marked improvement in the performance of the YOLOv4-tiny network. Following the application of the Mosaic data enhancement method and the incorporation of the multi-angle dataset, an increase in the mean Average Precision (mAP) value was observed. The mAP value rose to 68.84\% on the test set, an increase of 1.24\% over the pre-optimization performance. This quantifiable improvement validates the effectiveness of the proposed optimization techniques and confirms the enhanced efficiency of the YOLOv4-tiny network in vehicle marker recognition in natural scenes.
|
|---|