Learning method and learning device for sensor fusion to integrate information acquired by radar capable of distance estimation and information acquired by camera to thereby improve neural network for supporting autonomous driving, and testing method and testing device using the same

A method for testing a Convolutional Neural Network (CNN) by using a camera and a radar together, to thereby allow the CNN to perform properly even when an object depiction ratio of a photographed image, which represents a probability of one or more objects in the photographed image being illustrated properly, acquired through the camera is low due to a bad condition of a photographing circumstance, comprising steps of:

• (a) on condition that (1) a learning device, if a multichannel integrated image for training, which is generated by using (i) a photographed image for training acquired through a camera for training on a subject vehicle for training interworking with the learning device and (ii) a depth image for training acquired through a radar for training thereon, is acquired, has instructed at least one convolutional layer in the CNN to apply at least one convolutional operation to the multichannel integrated image for training, to thereby generate at least one feature map for training, on which information on the photographed image for training and information on the depth image for training are reflected;
• (2) the learning device has instructed at least one output layer in the CNN to apply at least one output operation to the feature map for training, to thereby generate estimated object information for training on the objects for training in the multichannel integrated image for training;
• and (3) the learning device has instructed at least one loss layer in the CNN to generate at least one loss by using the estimated object information for training and Ground-Truth (GT) object information corresponding thereto, and to perform backpropagation by using the loss, to thereby learn at least part of parameters in the CNN, a testing device instructing the convolutional layer in the CNN to apply the convolutional operation to a multichannel integrated image for testing which is generated by using (i) a photographed image for testing acquired through a camera for testing on a subject vehicle for testing interworking with the testing device and (ii) a depth image for testing acquired through a radar for testing thereon, to thereby generate at least one feature map for testing, on which information on the photographed image for testing and information on the depth image for testing are reflected;
• and
• (b) the testing device instructing the output layer in the CNN to apply the output operation to the feature map for testing, to thereby generate estimated object information for testing on the objects for testing in the multichannel integrated image for testing;
• wherein, at the step of (a), the testing device (i) acquires information for testing on one or more distances for testing and one or more angles for testing of the objects for testing from the subject vehicle for testing by referring to the depth image for testing, (ii) finds one or more object coordinates for testing, corresponding to at least part of the objects for testing, on the photographed image for testing by referring to information for testing on the distances for testing and the angles for testing, (iii) generates at least one guide channel image for testing by setting values which are generated by referring to the object coordinates for testing and probability distributions for testing, as its pixel values for testing included in the guide channel image for testing, and (iv) generates the multichannel integrated image for testing by channel-wise concatenating the guide channel image for testing with the photographed image for testing.

A learning device for training a Convolutional Neural Network (CNN) by using a camera and a radar together, to thereby allow the CNN to perform properly even when an object depiction ratio of a photographed image, which represents a probability of one or more objects in the photographed image being illustrated properly, acquired through the camera is low due to a bad condition of a photographing circumstance, comprising:

• at least one memory that stores instructions;
• and
• at least one processor configured to execute the instructions to perform processes of:
• (I) if a multichannel integrated image, which is generated by using (i) the photographed image acquired through the camera on a subject vehicle interworking with the learning device and (ii) a depth image acquired through the radar thereon, is acquired, instructing at least one convolutional layer in the CNN to apply at least one convolutional operation to the multichannel integrated image, to thereby generate at least one feature map, on which information on the photographed image and information on the depth image are reflected;
• (II) instructing at least one output layer in the CNN to apply at least one output operation to the feature map, to thereby generate estimated object information on the objects in the multichannel integrated image;
• and
• (III) instructing at least one loss layer in the CNN to generate at least one loss by using the estimated object information and Ground-Truth (GT) object information corresponding thereto, and to perform backpropagation by using the loss, to thereby learn at least part of parameters in the CNN;
• wherein, at the process of (I), the processor (i) acquires information on one or more distances and one or more angles of the objects from the subject vehicle by referring to the depth image, (ii) finds one or more object coordinates, corresponding to at least part of the objects, on the photographed image by referring to information on the distances and the angles, (iii) generates at least one guide channel image by setting values which are generated by referring to the object coordinates and probability distributions, as its pixel values included in the guide channel image, and (iv) generates the multichannel integrated image by channel-wise concatenating the guide channel image with the photographed image.

A testing device for testing a Convolutional Neural Network (CNN) by using a camera and a radar together, to thereby allow the CNN to perform properly even when an object depiction ratio of a photographed image, which represents a probability of one or more objects in the photographed image being illustrated properly, acquired through the camera is low due to a bad condition of a photographing circumstance, comprising:

• at least one memory that stores instructions;
• and
• at least one processor configured to execute the instructions to perform processes of:
• (I) on condition that (1) a learning device, if a multichannel integrated image for training, which is generated by using (i) a photographed image for training acquired through a camera for training on a subject vehicle for training interworking with the learning device and (ii) a depth image for training acquired through a radar for training thereon, is acquired, has instructed at least one convolutional layer in the CNN to apply at least one convolutional operation to the multichannel integrated image for training, to thereby generate at least one feature map for training, on which information on the photographed image for training and information on the depth image for training are reflected;
• (2) the learning device has instructed at least one output layer in the CNN to apply at least one output operation to the feature map for training, to thereby generate estimated object information for training on the objects for training in the multichannel integrated image for training;
• and (3) the learning device has instructed at least one loss layer in the CNN to generate at least one loss by using the estimated object information for training and Ground-Truth (GT) object information corresponding thereto, and to perform backpropagation by using the loss, to thereby learn at least part of parameters in the CNN, instructing the convolutional layer in the CNN to apply the convolutional operation to the multichannel integrated image for testing which is generated by using (i) a photographed image for testing acquired through a camera for testing on a subject vehicle for testing interworking with the testing device and (ii) a depth image for testing acquired through a radar for testing thereon, to thereby generate at least one feature map for testing, on which information on the photographed image for testing and information on the depth image for testing are reflected;
• and
• (II) instructing the output layer in the CNN to apply the output operation to the feature map for testing, to thereby generate estimated object information for testing on the objects for testing in the multichannel integrated image for testing;
• wherein, at the process of (I), the processor (i) acquires information for testing on one or more distances for testing and one or more angles for testing of the objects for testing from the subject vehicle for testing by referring to the depth image for testing, (ii) finds one or more object coordinates for testing, corresponding to at least part of the objects for testing, on the photographed image for testing by referring to information for testing on the distances for testing and the angles for testing, (iii) generates at least one guide channel image for testing by setting values which are generated by referring to the object coordinates for testing and probability distributions for testing, as its pixel values for testing included in the guide channel image for testing, and (iv) generates the multichannel integrated image for testing by channel-wise concatenating the guide channel image for testing with the photographed image for testing.