Machine learning based automated object recognition for unmanned autonomous vehicles
A platform is positioned within an environment. The platform includes an image capture system connected to a controller implementing a neural network. The neural network is trained to associate visual features within the environment with a target object utilizing a known set of input data examples and labels. The image capture system captures input images from the environment and the neural network recognizes features of one or more of the input images that at least partially match one or more of the visual features within the environment associated with the target object. The input images that contain the visual features within the environment that at least partially match the target object are labeled, a geospatial position of the target object is determined based upon pixels within the labeled input images, and a class activation map is generated, which is then communicated to a supervisory system for action.
1. A method comprising the steps of:
positioning a platform within an underwater environment, the platform including a sonar capture system connected to a controller, the controller having a neural network implemented therein, the neural network trained to associate one or more visual features within the underwater environment with a target object of interest utilizing a known set of input data examples and labels using the following operations:
1) Save a target example sub-image from input located around an object coordinates;
2) Select other sub-images from the input that do not contain the object coordinates;
3) Save the other sub-images as not-target sub-images;
4) Organize a dataset to label the target example sub-image with a target label and the not-target sub-images with a not-target label;
capturing, using the sonar capture system, one or more input images from the underwater environment;
recognizing, using calculated output values of at least one convolutional filter of the neural network, features of the one or more input images that at least partially match the one or more of the visual features within the underwater environment with the target object of interest;
calculating, using a classifier of the neural network, a label vector for the one or more input images where the label vector is a confidence value associated with the target object of interest;
labeling, using the confidence value from the classifier of the neural network, the one or more input images that at least partially match the one or more visual features within the underwater environment that at least partially match the target object of interest;
locating, using a combined label vector and class activation map output of the neural network, pixels within the one or more input images that at least partially match one or more visual features within the underwater environment that at least partially match the target object of interest; and
locating, using the controller, a geospatial position of the target object of interest based upon the pixels within the one or more input images.
2. The method of claim 1 , wherein the confidence value is determined to categorize the input image as containing the target object of interest.
3. The method of claim 1 further comprising the step of assigning a label to each pixel of the input image that represents the target object of interest.
4. The method of claim 3 , further comprising the step of using the controller to generate a class activation map of the input image using the labels assigned to each pixel of the input image.
5. The method of claim 4 further comprising the step of communicating the class activation map to a supervisory system networked with the platform.
6. The method of claim 4 further comprising the steps of:
processing values of the class activation map to yield modified values that reflect a goal of the controller, wherein the processing includes normalizing and thresholding the values of the class activation map to provide the geospatial position of the target object of interest; and
communicating the processed values of the class activation map to a supervisory system networked with the platform.
7. The method of claim 4 , wherein the class activation map has a heat map color overlay on the input image.
8. The method of claim 1 , wherein the step of locating the geospatial position of the target object of interest comprises translating coordinates of pixels with a target object label to coordinates representing a geospatial location.
9. The method of claim 1 further comprising the step of using the geospatial position of the target object of interest to guide decisions of the controller.
10. The method of claim 1 further comprising the step of transmitting the geospatial position of the target object of interest to a supervisory system networked with the platform.
11. A system comprising:
a platform positioned within an underwater environment, the platform including a sonar capture system connected to a controller, the controller having a neural network implemented therein, the neural network trained to associate one or more visual features within the underwater environment with a target object of interest utilizing a known set of input data examples and labels using the following operations:
1) Save a target example sub-image from input located around an object coordinates;
2) Select other sub-images from the input that do not contain the object coordinates;
3) Save the other sub-images as not-target sub-images;
4) Organize a dataset to label the target example sub-image with a target label and the not-target sub-images with a not-target label;
wherein the controller is configured to cause the sonar capture system to capture one or more input images from the underwater environment;
wherein the controller is configured to use calculated output values of at least one convolutional filter of the neural network to recognize features of one or more of the input images that at least partially match one or more of the visual features within the underwater environment with the target object of interest;
wherein the controller is configured to use a classifier of the neural network to calculate a label vector for the one or more input images where the label vector is a confidence value associated with the target object of interest;
wherein the controller is configured to label, using the confidence value from the classifier of the neural network, the one or more input images that at least partially match the one or more visual features within the underwater environment that at least partially match the target object of interest;
wherein the controller is configured to locate, using a combined label vector and class activation map output of the neural network, pixels within the one or more input images that at least partially match the one or more visual features within the underwater environment that at least partially match the target object of interest; and
wherein the controller is configured to locate a geospatial position of the target object of interest based upon the pixels within the one or more input images.