Robustness of machine learning pedestrian signal detection applied to pedestrian guidance device for persons with visual impairment

Research objective is to develop a robust image processing based pedestrian signal detection system for assisting independent walking of person with impaired vision. The developed system applies machine learning to enhance resistance to environmental disturbances: weather, season, or daylight. A hand-mounted guidance device is assembled with USB Web camera, vibration motor, Braille cell and portable computer. The acquired images are processed to discriminate the timing of safe road crossing, and information is transmitted by Braille and vibration. The system applied machine learning with Haar-like features and evaluated the effect of feeding the images that have disturbance of daylight and various luminance to improve the detection accuracy of the algorithm. Videos were recorded on total 24 conditions under sidewalk walking. The recognition result indicate that the pedestrian signals were correctly detected in 93.09% of the total image frames. The effect of the learning method showed robustness to disturbance such as lighting, caused by the validation of daytime and weather.