In this paper, we proposed an object detection method that combined a parallel network and a candidate region estimation process by using the SSD (Single Shot multibox Detector), which was a real-time object tracking algorithm based on deep learning. The proposed method used VGG-16 networks to extract feature maps in the convolution layer and tracked the objects through candidate region estimation. We used convolution layers containing semantic information and convolution layers containing detailed information. In order to use the information on the two layers together, the network structures proceeded with parallel. After extracting the feature map, the number of candidate bounding boxes was added and the candidate bounding box was used to generate the final bounding box. Conventional SSD may fail to detect objects that have a single structure and contain a small amount of information on the feature map. The proposed method used feature maps that contained more information to improve the object tracking rate. The proposed method showed a 2.2% improvement in mAP (mean average precision) values when compared with the conventional SSD method.
Volume 11 | 07-Special Issue
Pages: 1773-1780