Exemplar-based Object Detection using Car-mounted Fisheye Cameras for 360° Object Detection

摘要

This paper describes a practical approach to detect objects around a vehicle. We use fisheye cameras to monitor the area surrounding the vehicle. Fisheye cameras provide wide FOV at the cost of strong distortions. Therefore, one cannot apply common object detectors without large loss of accuracy. In order to deal with it, we introduce a panorama conversion that minimizes the distortion while maintaining the original wide FOV. Moreover, our conversion cancels the differences of both intrinsic and extrinsic parameters, and makes it possible to use the same training result for various cameras. Although the distortions are attenuated in the panorama images, there are some residual distortions. They result in large inner-class variance. We use a recently proposed exemplar-based object detector, Exemplar Network, to overcome it. We present two methods to balance the computational cost and detection accuracy of EN, then evaluate them with the data collected in public roads.