Detection of vague object signatures on deep learning surveillance devices

Abstract

The deep learning of object detection has become a breakthrough in recent years. Many papers demonstrated that this method records significant reliability results. However, the question arises whether objects that were successfully detected are initially conditioned clear in daylight. The object being detected is in the form of a photographic product that has numerous problems. It can be distant or have low-contrast so that their signatures are challenging to recognize, especially detection of persons in surveillance systems for dark-environments. This paper contributes to proving the deep learning method capable of detecting night-person (NP) with high precision and recall in the dark without image enhancement, by using ordinary cameras which operate on day-night or visible-near infrared spectrum runs on embedded systems. For that, an infrared-cut filter mechanical shutter is designed to block for the day or deliver infrared light for the night. The NP signatures are illuminated by an external infrared light source, providing three-channel high-resolution images. The distance of a NP from the camera becomes a decisive successful detection. The external infrared light source makes objects under or overexposed affecting the object being recognized. The validation with thoroughly new data of the NP constantly provides high precision and recall.