300 likes | 313 Views
Develop reliable methods of detecting, localizing, and classifying features associated with curbs using in-vehicle, low-cost, monocular vision sensor. Localize curbs within a range of 5 meters with 99% accuracy.
E N D
Objectives Develop reliable methods of detecting, localizing, and classifying featuresassociated with curbs using in-vehicle, low-cost, monocular vision sensor Localize curbs within a range of 5 meters with 99% accuracy
Approaches for Curb Detection Using Mono Camera Images • Appearance-based image analysis (~ Nov. 2013) • Extract features • Evaluate performance • Geometry-based image analysis (~ May. 2014) • Structure-from-motion to estimate camera motion • Multi-resolution plane sweeping algorithm to create 3-D point cloud • Plane fitting to detect curb • Combine appearance and geometric analysis (This Review)
Approaches for Curb Detection Using Mono Camera Images • Appearance-based image analysis (~ Nov. 2013) • Extract features • Evaluate performance • Geometry-based image analysis (~ May. 2014) • Structure-from-motion to estimate camera motion • Multi-resolution plane sweeping algorithm to create 3-D point cloud • Plane fitting to detect curb • Combine appearance and geometric analysis (This Review)
Detect Curb Using HOG* Feature Curb model HOG image Input image * Histogram of Oriented Gradients
Approaches for Curb Detection Using Mono Camera Images • Appearance-based image analysis (~ Nov. 2013) • Extract features • Evaluate performance • Geometry-based image analysis (~ May. 2014) • Structure-from-motion to estimate camera motion • Multi-resolution plane sweeping algorithm to create 3-D point cloud • Plane fitting to detect curb • Combine appearance and geometric analysis (This Review)
Geometry-based image analysis Input image Depth image 3-D point cloud Ground plane estimation
Approaches for Curb Detection Using Mono Camera Images • Appearance-based image analysis (~ Nov. 2013) • Extract features • Evaluate performance • Geometry-based image analysis (~ May. 2014) • Structure-from-motion to estimate camera motion • Multi-resolution plane sweeping algorithm to create 3-D point cloud • Plane fitting to detect curb • Combine appearance and geometric analysis (This Review)
Schematic Overview Input at t Appearance at t Candidate regions Annotate curb region Input at t+1 Appearance at t+1 Geometry
Appearance • For each image, divide intom x n grids • m: image height /grid size (pixels) • n: image width / grid size (pixels) Image at t
Appearance • For each grid, classify among two classes (road, curb) • uniform Local Binary Pattern (LBP) Image at t
Local Binary Pattern threshold Binary: 00011110 Decimal: 30 255 0 1 2 3
Appearance Output at t Output at t+1 Intersect • Once all the grids of two images are classified, get the intersection of them
Geometry • Green lines shows the vectors from the interesting points of image at time t (blue dots) to those of image at time t+1 (red dots) • Calculate the 3-D points using camera matrix
Appearance + Geometry • For each grid, • Fit the best plane using 3-D points • Compute the normal vector • Determine the normal vector is a road surface or a curb surface
Extend Curb Region • If the appearances are similar, extend the curb region • Calculate the distance of LBPs using chi-square
Track Curb Region Input at t+2 Input at t+3 Input at t+4 • For the next frames, tracking the appearance of the curbs • When tracking, keep checking the geometry constraint to remove the false positives if exist
Curved curb case Combine Analyses Extend Curb Region
Curb Detection using Production Camera Image size : 480 by 640 FOV: 180 degree
Test Curb Detection Image Size: 640 x 480 (pixels) ROI: 640 x 160 (pixels) Size of grid: 20 x 20 (pixels) Number of grids: 32 x 8 Output of the appearance-based curb detection
Test Curb Detection Remove outliers based on cluster size Find edges using Canny operator inside candidate region
Test Curb Detection - Fit polynomials to each segments, and check lines for similar curvatures (blue), and remove high curvatures (red) Annotate curb region on the original input image
Thank you Questions ?