Optimization of a CNN-based Object Detector for Fisheye Cameras

Fisheye cameras are widely used in the automobile industry, due to their wide field of view for the environment. There are a lot of algorithms that researchers use for object detection, among them Convolutional Neural Networks (CNN) detectors have been widespread during the course of the last decade, however, they are mostly trained and applied on conventional pinhole camera images. In this paper, an attempt to optimize a CNN-based detector for fisheye cameras was made, taking into consideration the barrel distortion, which complicates the object detection. Several experiments were carried out to optimize one state-of-the-art detector, using synthetic fisheye effect on training dataset. The obtained result proves that fisheye augmentation can considerably advance a CNN-based detector’s performance on fisheye images in spite of the distortion. Along the way, a new object detection framework based on the Tensorflow Estomator API was constructed to perform the experiments as convenient as possible.