Overview
This high-speed computer vision library for Arm NEON based processors was developed by Uncanny Vision (Head Office: India).
With changes to hardware completely unnecessary, systems can be made faster simply by replacing the currently used functions with the functions in this library.
Features
• Super high-speed C language functions for computer vision
• Optimized for Arm NEON (for Arm Cortex-A8/A9/A15/A7/A5 core processors based on Arm NEON)
• Performance is 2 to 20 times faster than non-optimized code such as OpenCV
• More than 70 function types (with more under development)
• A free evaluation version is provided
Application fields
• ADAS (Advanced Driver Assistance Systems)
• Machine vision
• Surveillance cameras
• Computer games
• AR (Augmented Reality)
• Gesture recognition
• Camera-based smartphone apps, wearable computing
• Robots
In addition, various other embedded computer vision systems
Optimized operations
• Effective utilization of Arm NEON
• Floating points rewritten to fixed points
• Data minimization
• Use of look-up tables
• Elimination of data transfer
• Data reuse
• Use of integral image
• Optimization at the algorithm level
• Effective utilization of cache
• Calculations based on fixed points
Benchmark examples
Algorithm | Throughput (megapixel/second) | Speed increase compared to non-optimized code, such as OpenCV |
Canny Edge detection
| 25
| 3 times
|
ORB (1500 key points)
| 3.7
| 5 times
|
Convolution filter 5x5
| 96
| 22 times |
Dilate / Erosion | 153 | 6.5 times |
Integral Image filter | 96 | 2.4 times |
Harris Corner Detection | 15.7 | 6.5 times |
Fast9 Key point detection | 24 | 2 times |
Face detection (LBP cascade) | - | 3.5 times
|
Connected Components | (Image Dependent)
| 1.7 times |
Pedestrian detection using HOG | 1.7 (on Arm Cortex-A15) | 9 times |
*Measured on Arm Cortex-A9 1 GHz. (On Arm Cortex-A15 800 MHz for HOG only)
Examples of use in automotive systems
• Pedestrian Detection
• Object Detection using HOG and LBP cascade detectors
• Forward Collision Warning
• Stereo Disparity based Warning system
• Lane Departure Warning
• Car Reverse Warning
• Surround View Systems
• Blindspot detection
Algorithm examples
Pedestrian Detection |
Car Detection |
 |
 |
Stereo Disparity |
Lane Detection |
 |
 |
Face Detection
|
 |
Algorithm list
Algorithms - High Level
Stereo Disparity
Pedestrian detection using HOG
Face Detection using LBP
Lane departure
Vehicle Detection
ORB
Lucas Kanade optical flow
Background Substraction
Tamper Detection
Stereo Disparity Post Processing
Dense Optical Flow
Face Post Estimation
Algorithms - Mid Level
Hough line detection
Lens Distortion Correction
Perspective transform
Connected Components
Integral Image
Non-Maximal Suppression
Kmeans clustering, based image segmentation
Harris Corner detection
Canny edge detection
Fast9 and Fast12 corner detection
Homography estimation
Algorithms - Low Level
Convolution kernels for different data types
Morphological operations - erosion, dilation
Image resizing
Histogram
Pyramid - Averaging, Gaussian
Array Multiplication
Sobel edge detection
Color conversion - RGB2YUV, RGB2HSV, RGB2LAB
Flip, Transpose
Table Lookup
Rotation - 90, 180, 270
Related Products
 |
Uncanny Vision
Deep-learning library |
 |
Uncanny Vision
Deep-learning library evaluation kit |
 |
Uncanny Vision
Super high-speed computer vision library for Arm NEON based processors |