A Comparative Study on HSV-based and Deep Learning-based Object Detection Algorithms for Pedestrian Traffic Light Signal Recognition
Nazirah Hassan, Kong Wai Ming, Choo Keng Wah · 2020
Small object detection has been a challenge in many image analysis applications. One such application is the ability to detect the status of Pedestrian Traffic Light (PTL) signal to allow decisions to be made by an intelligent system. The challenge is becoming more complex due to the increased complexity of urban environment, where objects of close similarity would confuse the detection mechanism. In this research, a study is carried out to compare two methods for the detection of small objects within large-sized images. The first method is a classical color-based segmentation approach while the second uses an intricate Deep Learning (DL) object detection algorithm. In the classical approach, objects within the selected range of Hue, Saturation and Value (HSV) composition are identified and extracted from the large-sized images. For DL approach, a Mask R-CNN was used where traffic light-like objects are identified by object instance segmentation process. From this research, it is shown that a two-tier approach, a hybrid HSV -DL model can detect the PTL signal directly and accurately from large-sized images in real-time on smart devices at an accuracy of 92.75%.