The investigation into the use of YOLO on overlay images for the purpose of privacy protection

Abstract

Human Detection is an important issue from the development of autonomous cities. That technology is used not only by cars but also by CCTVs which use that in event prevention. One of the most significant issues of human detection is privacy safety. Most conventional methods allow not only for detection but also for the recognition of humans, which introduce big concern about privacy protection. This is reason why challenge ChaLearn LAP Challenge @FG2020 "Identity-preserving Human Detection (IPHD)" was initiated, which focus on providing more accurate solution for alternatives types of cameras, which provide better privacy than classic types of cameras. That can be achieved by using different types of of detectors such as You look only Once (YOLO), Region-based Convolutional Neural Networks (RCNN). As YOLO detectors is one of best performing all experiments are focused on improvi-ng detection based on different version of this detector family and determining the most promising version of YOLO for this use case. As provided datasets are synchronised, work also is investigating values for overlays, created by overlapping images. Thermal overlay is done by covering depth image with partially transparent layer made from corresponding thermal image. Similar process is done for depth overlays. Goals of project are to find best version of detector and narrow down possible ranges for overlay. Achievement of improvement within 4-10 percent points on each dataset using v4 and 10-18 percent points for v5 and finding narrowed parameters ranges to [0.35,0.70] for depth overlay and [0.70,0.85] for thermal overlay allows me to conclude that all aims of the project were accomplished.