Undefined Object Detection Based on Convolutional Neural Network Using Small Dataset

In this paper, we propose a CNN-based method for undefined object detection. This method utilizes a small dataset acquired by multiple cameras at different views. The first procedure is to divide an image into
$64\times 64$
patches, and then the patches with the annotations input to CNN architecture for training a classification model. The annotation is a binary value determined by the area ratio of a marked area to the patch. In the detection process, our method applies the trained model with two estimated thresholds to detect undefined objects in the monitored area. The experiment results demonstrate that Resnet18 outperforms the compared CNN models.