Improvement of Deep Learning-based Human Detection using Dynamic Thresholding for Intelligent Surveillance System

Human detection plays an important role in many applications of the intelligent surveillance system (ISS), such as person re-identification, human tracking, people counting, etc. On the other hand, the use of deep learning in human detection has provided excellent accuracy. Unfortunately, the deep-learning method is sometimes unable to detect objects that are too far from the camera. It is because the threshold selection for confidence value is statically determined at the decision stage. This paper proposes a new strategy for using dynamic thresholding based on geometry in the images. The proposed method is evaluated using the dataset we created. The experiment found that the use of dynamic thresholding provides an increase in F-measure of 0.11 while reducing false positives by 0.18. This shows that the proposed strategy effectively detects human objects, which is applied to the ISS.