Low power consumption deep neural network for simultaneous object detection and semantic segmentation in images on a mobile computing device
A mobile computing device receives an image from a camera physically located within a vehicle. The mobile computing device inputs the image into a convolutional model that generates a set of object detections and a set of segmented environment blocks in the image. The convolutional model includes subsets of encoding and decoding layers, as well as parameters associated with the layers. The convolutional model relates the image and parameters to the sets of object detections and segmented environment blocks. A server that stores object detections and segmented environment blocks is updated with the sets of object detections and segmented environment blocks detected in the image.
1. A method to be performed by a computing device comprising:
receiving an image from a camera;
inputting the image into a convolution model loaded into a memory of the computing device and configured to generate a set of object detections for objects appearing in the received image and a set of segmented environment blocks, the convolution model comprising:
a plurality of parameters associated with a plurality of layers of the convolution model, the layers comprising a subset of encoding layers and a subset of decoding layers, with at least one skip connection between the encoding layers and the decoding layers;
a function relating the image and the parameters to the set of object detections and the set of segmented environment blocks; and
providing the set of object detections and the set of segmented environment blocks to a server computing device.
2. The method of claim 1 , wherein the subset of decoding layers further comprises a set of paired detector block heads for generating object detections of a large size.
3. The method of claim 1 , wherein the subset of decoding layers further comprises a linknet block and a set of paired detector head blocks for generating object detections of a medium size.
4. The method of claim 1 , wherein the subset of decoding layers further comprises a linknet block and a set of paired detector head blocks for generating object detections of a small size.
5. The method of claim 1 , wherein the subset of decoding layers further comprises a linknet block and at least one up-sampling layer to generate the set of segmented environment blocks.
6. The method of claim 1 , wherein the objects identified in the set of object detections are physically located in the environment and are captured within the image by the camera, wherein:
the objects comprise discrete shapes around which bounding boxes can be placed,
the objects are represented by a number of pixels located in the image, and
the objects are one of a plurality of object types.
7. The method of claim 1 , wherein the segmented environment blocks identified in the set of segmented environment blocks are physically located in the environment and are captured within the image by the camera, wherein:
the segmented environment blocks comprise unbounded shapes, including those behind or crossing one or more of the objects,
the segmented environment blocks are represented by a number of pixels located in the image, and
the segmented environment blocks are one of a plurality of block types.
8. The method of claim 1 , wherein the parameters were obtained using a set of training images, each training image associated with a set of training object detections and a set of training segmented environment blocks.
9. The method of claim 1 , wherein the encoding layers each comprise at least one resnet block.
10. The method of claim 1 , wherein the decoding layers further comprise a linknet block, the linknet block comprising:
a first convolution layer;
an upsampling layer;
a second convolution layer; and
a third convolution layer.
11. The method of claim 1 , further comprising:
responsive to determining that one of the identified set of object detections adheres to a set of rules loaded into the memory of the computing device, providing a notification via the mobile computing device to a user.
12. The method of claim 11 , wherein the rule indicates that a road hazard has been detected within a proximity of a vehicle and wherein the notification informs the user of the road hazard.
13. The method of claim 11 , wherein the rule indicates that a waypoint or a destination has been detected within a proximity of a vehicle and wherein the notification informs a user of the presence of the waypoint or the destination.
14. The method of claim 1 , wherein the computing device is a mobile computing device.
15. A non-transitory computer-readable storage medium of a computing device storing instructions that, when executed by a computer processor of the computing device cause the computing device to:
input an image into a convolution model loaded into a memory of the computing device and configured to generate a set of object detections for objects appearing in the image and a set of segmented environment blocks, the convolution model comprising:
a plurality of parameters associated with a plurality of layers of the convolution model, the layers comprising a subset of encoding layers and a subset of decoding layers, with at least one skip connection between the encoding layers and the decoding layers;
a function relating the image and the parameters to the set of object detections and the set of segmented environment blocks; and
provide the set of object detections and the set of segmented environment blocks to a server computing device.
16. The computer-readable storage medium of claim 15 , wherein the subset of decoding layers further comprises a set of paired detector block heads for generating object detections of a large size.
17. The computer-readable storage medium of claim 15 , wherein the subset of decoding layers further comprises a linknet block and a set of paired detector head blocks for generating object detections of a medium size.
18. The computer-readable storage medium of claim 15 , wherein the subset of decoding layers further comprises a linknet block and a set of paired detector head blocks for generating object detections of a small size.
19. The computer-readable storage medium of claim 15 , wherein the subset of decoding layers further comprises a linknet block and at least one up-sampling layer to generate the set of segmented environment blocks.
20. The computer-readable storage medium of claim 15 , wherein the objects identified in the set of object detections are physically located in an environment and are captured within the image by a camera, wherein:
the objects comprise discrete shapes around which bounding boxes can be placed,
the objects are represented by a number of pixels located in the image, and
the objects are one of a plurality of object types.