Current Research

Pedestrian Detection / Tracking / Identification

I am currently working on Pedestrian-centric Image Processing for videos of various cameras such as surveillance cameras and in-vehicle cameras. Now I’m mainly tackling Pedestrian Detection from an in-vehicle camera. Especially, my main focus is detecting distant pedestrians who are very small in an image captured by such cameras.

For several years, I have been studying processing surveillance videos, including Pedestrian Tracking and Pedestrian Image Retrieval based on the Person Re-identification, which is a technique to match the same pedestrian pairs in two different surveillance videos that are captured by different cameras observing non-overlapping areas. I have been tackling multiple pedestrians tracking across many cameras and also tackling Pedestrian Image Retrieval, which is an image retrieval of a specific pedestrian from surveillance videos. Through the researches, I published a survey of Person Re-identification researches (in Japanese). I also released a public pedestrian dataset named Shinpuhkan 2014 dataset.

I was also working on Privacy Protection of pedestrians in a surveillance camera. The system, named “HENSHIN (Metamorphosing) Camera, ” detects pedestrians and estimates the background image, and draws pedestrian avatars on the estimated background image. The system never outputs pedestrian raw image even if the system fails the pedestrian detection (Link: Henshin (Metamorphosing) Camera).