METHOD FOR DETECTING PSEUDO-3D BOUNDING BOX BASED ON CNN CAPABLE OF CONVERTING MODES ACCORDING TO POSES OF OBJECTS USING INSTANCE SEGMENTATION AND DEVICE USING THE SAME

摘要

The present invention relates to a method for detecting a CNN-based Pseudo-3D bounding box capable of switching a mode according to the posture of an object detected using instance segmentation, and according to this method, a pseudo-3D The shading information for each surface of the bounding box can be reflected in learning, the capital-3D bounding box is acquired through Lidar or radar, the surface can be segmented using a camera, and the detection method is The learning device causes the pooling layer to apply the pooling operation on the 2D bounding box to generate the pooled feature map, the FC layer to apply the neural network operation, and the convolutional layer to apply the convolution operation to the surface area. and causing the FC layer to generate class loss and regression loss.