(Jan 2023)Automatic Landmark Object Extraction and Image-based Visibility Estimation System, Prof. Wai-Lun LO (Principal Investigator), Dr. HSUNG Tai-Chiu (Co-I), Dr. Hong FU (Co-I, EDUHK), UGC/FDS13/E01/23, 24 months, $768,185
Project Summary
Visibility is an important safety indicator for road traffic, as under bright background conditions, an object on the ground can only be detected or recognised if it is within the visibility distance threshold. Visibility is also used as an environmental parameter for pollution and weather conditions’ monitoring.
In past research, we extracted image features and used meteorological laws for visibility estimation calculations. However, the accuracy of these approaches depends on image quality and is affected by noise. It is also difficult to extract all of the environmental factors via such approaches and formulate an equation relating these factors to visibility. Thus, artificial intelligence (AI) approaches are currently used for visibility estimation.
In preliminary research on AI-based visibility estimation, we used a pre-trained convolutional neural network to extract image features from webcam images. Instead of using whole digital images, effective areas were identified, and a single correlation variable was used to select suitable image features. A generalised regression neural network (GRNN) and a deep learning algorithm were designed and used to estimate mapping between visibility and selected images, with use of a visibility database provided by the Hong Kong Observatory (HKO). The results showed that this method is more accurate than the classical method based on handcrafted features. However, this method is not fully automated, as effective regions of image data are selected manually using expert judgement. Furthermore, the accuracy of this method is limited, as image features are selected using only a single correlation coefficient and low-resolution webcam images. Finally, the data training and application scope of this method are limited by the HKO-provided database. Thus, the developed algorithm needs to be tested with different datasets.
In this proposed project, we will investigate and develop an automatic landmark object (LMO) extraction system that can extract LMO regions from weather photos and identify the LMOs’ effective visibility estimation ranges. The mapping between LMO’s regions’ features values and visibility ranges for different ranges are approximated by an artificial neural network (ANN). A multi-class ANN visibility estimator will be developed in this project. We will also build a hardware prototype of the multi-class ANN Visibility Estimation System have the following functions: (i) automatic visibility meter data collection; (ii) automatic weather photos, image data collection; and (iii) wireless communication with a host computer (e.g. via Wi-Fi). Finally, the multi-class estimation algorithms will be implemented in a host computer. The team of this proposed project will evaluate the LMO extraction system and the Visibility Estimation and Training System by conducting experimental studies. The major outcome of the project is a low-cost and automatic visibility estimation system which contribute to environmental monitoring technology.
