Techniques for precisely locating landmarks in monocular camera images with deep learning
Advanced driver assistance (ADAS) systems and methods for a vehicle comprise capturing an image using a monocular camera system of the vehicle and detecting a landmark in the image using a deep neural network (DNN) trained with labeled training data including generating a bounding box for the detected landmark, predicting a depth of pixels in the image using a convolutional neural network (CNN) trained with unlabeled training data captured by a stereo camera system, filtering noise by averaging predicted pixel depths for pixels in the region of the bounding box to obtain an average depth value for the detected landmark, determining a coordinate position of the detected landmark using its average depth value, and performing at least one ADAS feature using the determined coordinate position of the detected landmark.
1. An advanced driver assistance system (ADAS) for a vehicle, the ADAS comprising:
a monocular camera system configured to capture an image; and
a controller configured to:
receive the image;
detect a landmark in the image using a deep neural network (DNN) trained with labeled training data including generating a bounding box for the detected landmark;
predict a depth of pixels in the image using a convolutional neural network (CNN) trained with unlabeled training data captured by a stereo camera system;
filter noise by averaging predicted pixel depths for pixels in the region of the bounding box to obtain an average depth value for the detected landmark;
determine a coordinate position of the detected landmark using its average depth value; and
perform at least one ADAS feature using the determined coordinate position of the detected landmark.
2. The ADAS of claim 1 , wherein the determined coordinate position of the detected landmark is a polar coordinate position comprising the average depth value and an angle from the monocular camera system to a center of the detected landmark.
3. The ADAS of claim 2 , wherein the ADAS feature comprises localizing a position of the vehicle on a global positioning system (GPS) map to the determined polar coordinate position.
4. The ADAS of claim 1 , wherein the ADAS feature comprises collision avoidance during an autonomous driving mode.
5. The ADAS of claim 1 , wherein the averaging of the predicted pixel depths in the region of the bounding box comprises applying an averaging filter.
6. The ADAS of claim 1 , wherein the CNN is configured to predict a depth of every pixel of the image.
7. The ADAS of claim 1 , wherein the unlabeled training data comprises pairs of simultaneously captured left and right images.
8. The ADAS of claim 1 , wherein the controller is further configured to perform vehicle-to-vehicle transfer learning to further train the DNN.
9. The ADAS of claim 1 , wherein the labeled training data comprises a plurality of images of different types of landmarks that are manually labeled by a human annotator.
10. A method of detecting a landmark from an image captured by a monocular camera system of a vehicle, the method comprising:
capturing, by the monocular camera system, the image;
receiving, by a controller of the vehicle, the image;
detecting, by the controller, a landmark in the image using a deep neural network (DNN) trained with labeled training data including generating a bounding box for the detected landmark;
predicting, by the controller, a depth of pixels in the image using a convolutional neural network (CNN) trained with unlabeled training data captured by a stereo camera system;
filtering, by the controller, noise by averaging predicted pixel depths for pixels in the region of the bounding box to obtain an average depth value for the detected landmark;
determining, by the controller, a coordinate position of the detected landmark using its average depth value; and
performing, by the controller, at least one advanced driver assistance system (ADAS) feature using the determined coordinate position of the detected landmark.
11. The method of claim 10 , wherein the determined coordinate position of the detected landmark is a polar coordinate position comprising the average depth value and an angle from the monocular camera system to a center of the detected landmark.
12. The method of claim 11 , wherein the ADAS feature comprises localizing a position of the vehicle on a global positioning system (GPS) map to the determined polar coordinate position.
13. The method of claim 10 , wherein the ADAS feature comprises collision avoidance during an autonomous driving mode.
14. The method of claim 10 , wherein the averaging of the predicted pixel depths in the region of the bounding box comprises applying an averaging filter.
15. The method of claim 10 , wherein the CNN is configured to predict a depth of every pixel of the image.
16. The method of claim 10 , wherein the unlabeled training data comprises pairs of simultaneously captured left and right images.
17. The method of claim 10 , further comprising performing, by the controller, vehicle-to-vehicle transfer learning to further train the DNN.
18. The method of claim 10 , wherein the labeled training data comprises a plurality of images of different types of landmarks that are manually labeled by a human annotator.