Analysis of Image Thresholding Algorithms for Automated Machine Learning Training Data Generation

Abstract

Secured compounds often safeguard physical layout details of both internal and external facilities, but these details are at risk due to the growing inclusion of Light Detection and Ranging (LiDAR) sensors in consumer off-the-shelf (COTS) technology such as cell phones. The ability to record detailed distance data with cell phones facilitates the production of high-quality three-dimensional scans in a discrete manner which directly threatens the security of private compounds. Therefore, it behooves the organizations in charge of private compounds to detect LiDAR activity. Many security cameras already detect LiDAR sources as generic light sources in specific conditions, but further analysis must identify these light sources as LiDAR sources in order to alert an organization of a potential security incident. Testing confirms the feasibility of identifying some LiDAR sources based on the color and intensity of light shined directly into a camera sensor, but this analysis proves inadequate for cell phone LiDAR. However, the unique intensity and pattern characteristics of cell phone LiDAR reflected off a surface can potentially be identified by an object identification machine learning model. In order to train a model to identify a LiDAR object, we must first produce a training dataset containing marked and labelled LiDAR objects. To do this, we apply an image thresholding algorithm to isolate the LiDAR object in an image to calculate its bounding box. The image thresholding algorithm directly affects the bounding box accuracy, so we test two different algorithms and find that Otsu’s image thresholding algorithm performs best, resulting in 99.5% accurate bounding boxes.