부산대학교 도서관

This paper proposes to detect, recognize, and track a specific object in video sequences using Convolutional Neural Networks (CNNs). The CNNs used are the TensorFlow SSD model and Inception model, with a use case of airplane detection as a test subject, although it can be widely extensible to any class of objects as per the application. For the SSD model, images of planes were downloaded and annotated using bounding boxes to identify regions of interest. Training and test sets were split, after which TensorFlow specific records were generated. Whereas, for the Inception model, the last layer of the Neural Network was trained with multiple images of airplanes and random images to obtain a classifier for identify planes vs. no planes. The SSD model was accurate, generating crisp bounding boxes with a relatively high accuracy. The Inception model had a higher accuracy than the SSD model in terms of false positives and false negatives. But it does not display bounding boxes as the model is not meant to find the region of interest. The GPU outperformed the CPU in training and testing by a wide margin. The Inception model is suitable to extract frames in which a specific object is present if the position of the object is not of importance. The SSD model is suitable if the specific object needs to be detected with its position in a video frame. Keywords: TensorFlow, Object Recognition, Bounding Boxes, CNN, Neural Network, SSD, Inception, Plane, CPU, GPU
1. INTRODUCTION Object recognition is used in computer vision, video analytics and image processing tasks, wherein the instances of semantic objects of a certain class (such as humans, buildings, cars [...]