| Summary: | 碩士 === 國立屏東大學 === 資訊科學系碩士班 === 107 === This study proposes a tracking system that uses deep learning to pre-identify the classification of objects, and then based on the classified information to effectively track objects, which can continue to track after different objects are interlaced or briefly disappeared into the picture.
The system first identifies the objects in the picture through the YOLO neural network, obtains the classification and position information on the scene, and temporarily stores it in the identification table.Then, according to the classification information in the identification table, analyze the center points of each object and give them a number to track them, and the same numbered object indicates the same individual in the tracking, and the tracking information is updated to the tracking table. And the confidence value is used to keep the objects that have disappeared briefly from the picture so that they can be re-tracked with the same number.
The experimental results show that combined with the YOLO neural network and using the classification information and the central point location information, a fast and effective tracking effect can be achieved, and the multi-object on the scene can be simultaneously tracked at a computing speed of 45 frames per second and is not limited to a specific classification such as a car or a pedestrian. The system can be applied to input images of different resolutions, even if the resolution is 1440×1080, the operation speed will not be lower than 40 frames per second.
|
|---|
