Benchmark of Deep Learning Visual and Far-Infrared Videos Toward Weather-tolerant Pedestrian Traffic Monitoring

Sidewalks should be developed and integrated into city planning for economic growth and livability. And it is necessary to obtain data on pedestrians. One type of pedestrian data, pedestrian counting, commonly uses visible and far-infrared cameras to detect and track people. However, it is not clear in which situations the number of pedestrians can be efficiently estimated from each video. In this study, we use visible and far-infrared videos captured on sidewalks to estimate the number of pedestrians. And we investigate the effects of various times and weather on the estimation accuracy. Experimental results showed that the head count accuracy decreased when using visible video at night and in the rain. When using far-infrared video, the accuracy of head count decreased in the range of 23°C-30°C of ambient temperature. This study revealed that in order to achieve a highly accurate head count throughout the day, it is necessary to change the video used depending on the time and ambient temperature or to use both types of video.