| （英） |
A minimum sensor configuration is desired for a popular automatic vehicle. In this study, an omnidirectional camera with a wide viewing angle is used for object detection to recognize the driving environment. Since an omnidirectional camera has a different appearance from that of a normal monocular camera, geometric transformation is used to perform deep learning. However, the image is segmented at both ends of the object. In this study, we consider the image to be repeated and set annotations to detect objects that are segmented. In addition, since both ends of the omnidirectional image are far away from each other, it is difficult to extract features only by convolutional neural network, so Self Attention is added to the backbone. The accuracy of the network with Self Attention was measured by mAP, and the maximum mAP of the network with Self Attention was 66.5%, which made it possible to detect segmented objects. |