Outdoor Human Detection with Stereo Omnidirectional Cameras

Abstract

An omnidirectional camera can simultaneously capture all-round (360°) environmental information as well as the azimuth angle of a target object or person. By configuring a stereo camera set with two omnidirectional cameras, we can easily determine the azimuth angle of a target object or person per camera on the image information captured by the left and right cameras. A target person in an image can be localized by using a region-based convolutional neural network and the distance measured by the parallax in the combined azimuth angles.

Research on Omnidirectional Stereo Measurement Using Convex Mirrors and Vertical Disparity

“…In this experiment, the same experiment as in the previous study [ 10 ] was conducted for comparison with the previous study [ 10 ]. Target persons were photographed at 1.0 m intervals in the distance range of 1.0 m to 5.0 m from the omnidirectional stereo camera.…”

“…Compared to the previous study [ 6 ], omnidirectional measurement is now possible with half the number of cameras. Moreover, compared to the previous studies [ 7 , 8 , 9 , 10 ], the new system enables complete omnidirectional measurement without any blind spots. Furthermore, the accuracy was improved by 37.4% compared to the previous study [ 10 ], which had the best accuracy among the studies that enabled measurement with two cameras.…”

“…Moreover, compared to the previous studies [ 7 , 8 , 9 , 10 ], the new system enables complete omnidirectional measurement without any blind spots. Furthermore, the accuracy was improved by 37.4% compared to the previous study [ 10 ], which had the best accuracy among the studies that enabled measurement with two cameras. We consider this result to be a remarkable achievement in stereo measurement.…”

“…In the study by Tanaka Shunya et al [ 10 ], a commercially available all-sky camera [ 11 ] was utilized to perform object detection through machine learning and azimuth angle calculation to achieve omnidirectional measurement using two cameras. However, this method faced a drawback as it resulted in a blind spot in the camera’s extended baseline ( Figure 1 ) that could not be detected, so it was not a perfect omnidirectional measurement.…”

“…In the measurement experiment, the results were compared with those of the conventional method to verify the accuracy. The conventional method compared was Tanaka’s method [ 10 ], which was selected as the one with the best measurement accuracy among the conventional studies that have performed measurements under similar conditions. While the conventional method had a relative error of up to 40%, the new method succeeded in reducing the error to a maximum of 2.6%.…”