| Sonstiges: |
- Nachgewiesen in: USPTO Patent Applications
- Sprachen: English
- Document Number: 20230258471
- Publication Date: August 17, 2023
- Appl. No: 17/674596
- Application Filed: February 17, 2022
- Claim: 1. An AI based change detection system for executing a method to detect changes in geo-tagged videos to update HD maps, the method employing a neural network of modular components at least comprising a keyframe extraction module for processing two or more videos relating to separate traversals of an area of interest to which the HD map which is to be updated relates, a deep neural network module processing output of the keyframe extraction module, a change detection module processing output of the deep neural network module, and an auxiliary computations module which is designed to effectively aid the change detection module.
- Claim: 2. The AI based change detection system of claim 1, wherein the deep neural network module comprises a selection of network-modules for object classification, object detection, object tracking, depth estimation, semantic segmentation, and text recognition, respectively, wherein the modules are equipped to perform the following tasks a. Semantic Segmentation Network module: estimates per-pixel classification from images. b. Depth Estimation Network module: estimates the depth of objects from just 2D images. c. Object Detection Network module: performs localization and classification and outputs bounding box coordinates, and labels along with confidence scores. d. Object Tracking Network module: tracks an object through all the frames of the videos and outputs the bounding box coordinates. e. Object Classification Network module: classifies objects in the image and outputs the labels. f. Text Recognition Network module: performs text recognition.
- Claim: 3. The AI based change detection system of claim 1, wherein the keyframe extraction module is designed to extract frames from the videos with corresponding GPS information and to align and output a sequence of frames with corresponding GPS coordinates, and/or that when corresponding GPS information is missing, the keyframe extraction module interpolates the frames to convert them into geo-tagged frames.
- Claim: 4. The AI based change detection system of claim 1, wherein the auxiliary computations module is arranged to calculate vanishing points in the video frames enabling their use by the other modules of the system.
- Claim: 5. The AI based change detection system of claim 1, wherein the change detection module processes output of the deep neural network module, while simultaneously using output from the auxiliary computations module.
- Claim: 6. The AI based change detection system of claim 5, wherein the change detection module is arranged with submodules to detect changes in a. Traffic Signs b. Perpendicular Wall Barrier c. Hatched markings d. Text e. Arrow f. Overhead Structure g. Traffic Barrier h. Lane Boundary
- Claim: 7. The AI based change detection system of claim 6, wherein the submodule to detect whether traffic signs have been added or removed between different traversals represented by the videos to be processed comprises the tasks of object detection, tracking and classification taken from the deep neural network module, and a filtering block processing classification output of the deep neural network module to suppress false positives, followed by a framewise comparison to identify the changes in traffic signs.
- Claim: 8. The AI based change detection system of claim 2, wherein the change detection module processes output of the deep neural network module, while simultaneously using output from the auxiliary computations module.
- Claim: 9. The AI based change detection system of claim 8, wherein the change detection module is arranged with submodules to detect changes in a. Traffic Signs b. Perpendicular Wall Barrier c. Hatched markings d. Text e. Arrow f. Overhead Structure g. Traffic Barrier h. Lane Boundary
- Claim: 10. The AI based change detection system of claim 9, wherein the submodule to detect changes in perpendicular wall barrier obtains unit surface normals from depth maps as provided by the depth estimation network module, which are judged to be perpendicular to the surface, which unit surface normals are masked with the output of the semantic segmentation network module and subsequently denoised and thresholded to provide an output of the perpendicular walls in each frame, wherein a change in such perpendicular wall barrier is extracted by comparing the detected perpendicular walls at all GPS locations.
- Claim: 11. The AI based change detection system of claim 9, wherein the hatched markings are hatched road markings and changes in hatched road markings that are used to separate traffic to ensure safety are provided by having the semantic segmentation network module segment the hatched markings from the images of different traversals pertaining to an area of interest, and forward these hatched road markings to an area based filtering algorithm, the output of which is compared for said different traversals to detect the changes in the hatched road markings.
- Claim: 12. The AI based change detection system of claim 9, wherein the text comprises text on signboards and changes in text on signboards are provided by having the semantic segmentation network module segment objects with texts from the images of different traversals pertaining to an area of interest, and forward these texts to a text recognizer to identify the textual content which is compared for said different traversals to detect the changes in the texts.
- Claim: 13. The AI based change detection system of claim 9, wherein detection of locations where arrow signs have been added or removed is provided by arranging that the semantic segmentation network module obtains masks for the arrow signs, wherein contours extracted from these masks are smoothed by a merge and prune algorithm, followed by a low pass filter across frames to remove spurious detections, wherein a timeseries signal marking the presence of arrows in each frame is generated for the frame sets of each traversal in the area of interest, and a Fourier transformation is applied for comparing a shift in the frame sets, wherein said shift is used to align the frames and compare the aligned frames for extraction of the changes.
- Claim: 14. The AI based change detection system of claim 9, wherein the overhead structures comprise the faces of tunnels or bridges and the presence or absence of the overhead structures is detected using the segmentation network module from the deep neural network module and the vanishing point calculation of the auxiliary computations module, wherein the segmentation mask in each frame is weighed using the vanishing point to get a weighted area representing a mass of the overhead structure, wherein a center of said mass is computed to estimate an approximate depth of the structure, and wherein a sequence of masses and centroids forming clusters around the actual structure and a timeseries signal are processed and compared to detect added or removed structures.
- Claim: 15. The AI based change detection system of claim 9, wherein the traffic barriers comprise guardrails and the locations of added or removed traffic barriers are detected using the semantic segmentation network module from the deep neural network module and the vanishing point calculation of the auxiliary computations module, wherein a mask is weighed in its frame using vanishing point calculation to get a weighted area representing the mass of the traffic barrier, whereafter this mass signal across frames is converted to correspond to a distance from the starting frame providing a timeseries signal which is processed and compared across traversals in the area of interest to find added or removed traffic barriers.
- Claim: 16. The AI based change detection system of claim 9, wherein the locations of added or removed lane boundaries and determination of the type of lane boundary is detected using the semantic segmentation network module from the deep neural network module and the vanishing point calculation of the auxiliary computations module, wherein each frame is independently reduced to a tuple describing lane markings detected on the road from left to right, and defining the types and order of lane markings without their actual locations in the frame.
- Claim: 17. The AI based change detection system of claim 16, wherein providing the tuple is executed by intersecting semantic segmentation masks for the lane boundary type with edges, followed by line detection, noise filtering, and clustering/averaging of lines to single detection per lane boundary.
- Current International Class: 01; 06; 01; 06; 06
|
