System and method for mapping vectorization and navigation system using convolutional neural network
An onboard navigation system for vehicles in GPS-denied environments using image-based mapping. A mapping vectorization and navigation system uses a convolutional neural network (CNN) to improve the speed of recognition, orientation, and navigation while avoiding the use of GPS/GNSS signals.
1 . A method for navigating an autonomous vehicle in a location without the use of real-time satellite data, the method comprising:
providing a reference vectored map of the location, in a storage medium onboard the autonomous vehicle;
capturing a raster image of a landscape within the location using a imaging system onboard the autonomous vehicle;
creating a filtered image by applying an edge detection filter to the raster image to detect edges and define initial vector borders within the raster image of the landscape;
processing, under control of a processor onboard the autonomous vehicle, the filtered image through a first convolutional neural network (CNN) to produce a first vectored map;
comparing, under control of the processor onboard the autonomous vehicle, the generated first vectored map with the reference vectored map to determine the current position of the autonomous vehicle within the location; and
changing the position of the autonomous vehicle based on the determined current position of the autonomous vehicle.
2 . The method of claim 1 , further comprising:
processing, by an FPGA control block onboard the autonomous vehicle, the raster image with a second CNN to create a second vectored map, wherein the second vectored map comprises details not found in the first vectored map.
3 . The method of claim 1 , further comprising pretraining the first CNN on datasets comprising different types of landscape.
4 . The method of claim 3 , wherein the pretraining includes using training landscape datasets comprising images of specific environments, including urban, rural, forested, or aquatic environments.
5 . The method of claim 1 , further comprising fine-tuning the second CNN using real-time captured images specific to the autonomous vehicle's operational environment.
6 . The method of claim 2 , wherein the FPGA control block is further configured to adjust processing priorities between the edge detection filter and convolutional neural networks based on real-time computational load and power availability onboard the autonomous vehicle.
7 . The method of claim 1 , where the processing the filtered image through the first convolutional neural network (CNN) to produce the first vectored map further comprises using corresponding detected sets of landscape in the filtered images.
8 . The method of claim 1 , wherein creating a filtered image by applying an edge detection filter to the raster image to detect edges and define initial vector borders within the raster image includes applying a Sobel filter to determine the borders between different types of landscape areas, and turning these borders into vectors, wherein each vector is defined by the coordinates of two points, forming straight lines without any curves, to represent borders within the first vectored map.
9 . The method of claim 1 , wherein comparing the first vectored map and the reference vectored map comprises calculating the distance between vectors in the first vectored map and the reference map by subtracting their respective coordinates, with a lower calculated distance indicating a higher similarity between vectors, and using a quantity of similar vectors to assess the overall similarity of the first and previously stored vectored maps.
10 . A system for navigating an autonomous vehicle in a location without using real-time satellite data, the system comprising:
an autonomous vehicle;
a processor onboard the autonomous vehicle;
an onboard memory operably coupled to the processor;
a nonvolatile storage medium for storing raster images and reference vectored maps of the location;
an optical system comprising a camera, configured for capturing raster images of a landscape within the location;
an image processing module comprising an edge detection filter and one or more convolutional neural networks (CNN), wherein the edge detection filter is configured for detecting edges and defining initial vector borders within the raster images and wherein the one or more CNNs are configured for producing a first vectored map with corresponding detected sets of landscape by processing the filtered images through one or more CNNs;
a control block, operably connected to the image processing module, configured for comparing the generated first vectored map with the previously stored reference vectored maps to determine a current position within the location; and
a transmitter, configured for sending the current position within the location to a destination.
11 . The system of claim 10 , wherein an onboard control block processes the raster images with a second CNN to create a second vectored map, wherein the second vectored map comprises details not found in the first vectored map.
12 . The system of claim 10 , wherein the optical system comprises an FPGA control block configured to process images through the edge detection filter and the first CNN and a second CNN.
13 . The system of claim 10 , wherein the first CNN has been pretrained on datasets comprising different types of landscape.
14 . The system of claim 13 , wherein the pretraining of the first CNN includes using training landscape datasets comprising images of specific environments, including urban, rural, forested, or aquatic environments.
15 . The system of claim 13 , wherein the second CNN has been fine-tuned using real-time captured images specific to the autonomous vehicle's operational environment and wherein fine-tuning is adjusted based on feedback from location determination accuracy assessments.
16 . A method for determining the position of an autonomous vehicle without the use of real-time satellite data satellite data, the method comprising:
providing a reference vectored map of a location, in a storage medium onboard the autonomous vehicle;
capturing a raster image of a landscape within the location using a imaging system onboard the autonomous vehicle;
creating a filtered image by applying an edge detection filter to the raster image to detect edges and define initial vector borders within the raster image of the landscape;
processing, under control of a processor onboard the autonomous vehicle, the filtered image through a first convolutional neural network (CNN) to produce a first vectored map;
comparing, under control of the processor onboard the autonomous vehicle, the generated first vectored map with the reference vectored map to determine the current position of the autonomous vehicle within the location.
17 . The method of claim 16 , further comprising transmitting the current position of the autonomous vehicle to a remote device.
18 . The method of claim 16 , further comprising storing the location of the autonomous vehicle to the storage medium.
19 . The method of claim 16 , wherein a Field-Programmable gate array (FPGA) control block adjusts processing priorities between the edge detection filter and convolutional neural networks based on real-time computational load and power availability onboard the autonomous vehicle.
20 . The method of claim 16 , wherein the processing the filtered image through the first convolutional neural network (CNN) to produce the first vectored map further comprises using corresponding detected sets of landscape in the filtered images.