Vision Analytics Toolbox, Blackfin
Vision Analytics Toolbox, Blackfin
The Vision Analytics Toolbox (VAT) library for the Blackfin processor provides a set of software modules that make use of primitives from the video content analysis toolbox (VCAT) and provides solutions to relatively complex problems like face detection, edge detection, shape detection, classification, and others.
Top
Overview
Features and Benefits
- Provides optimized algorithms for Canny edge detection, Hough transform for circles and transform for lines, random tree classifier, feature generation for speed limit detection, circle/ellipse detection, face detection and connected component analysis.
- Supports both cache and DMA modes.
- Provision to operate directly on the output of PVP for ADSP-BF609 processors.
- Code compatible for all Blackfin processors except ADSP-BF535.
- Fully re-entrant and multi-instancing capable.
Product Details
Blackfin Vision Analytics Toolbox (VAT) is designed to be a software toolbox for system integrators working on vision based applications. VAT is a set of software modules that make use of primitives from ADI Blackfin Image Processing Toolbox and provides solutions to relatively complex problems like face detection, edge detection, shape detection, classification etc. This is at a level that is lower than a full scale application and higher than image/video processing primitives. These modules can be used by applications such as traffic sign recognition, iris detection and defect detection in assembly line automation.This product was formerly called as ADAS Vision Analytics Toolbox (AVAT). This was renamed to Blackfin Vision Analytics Toolbox (VAT) to make the name consistent with the broad range of applications where this toolbox can be used.
Licensing
Each module supports the Analog Devices, Inc. (ADI) Blackfin or SHARC Processor family and is a licensed product that is available in object code format. Recipients must sign or accept a license agreement with ADI prior to being shipped or downloading the modules identified in the license agreement.
Performance Metrics
MIPS summary:
| Module | Code RAM (KiB) |
Data RAM (KiB) |
Constant Data Tables (KiB) |
AverageCycles/Pel | MIPS |
| Canny Edge Detector (PVP mode) |
10.05 |
42.53 |
0.007 |
14.76 |
136 |
| Hough Transform Circle | 6.43 | 41.12 |
0.19 | 11.43 | 105 |
| Classifier | 0.93 | 0.048 | 0 | 296360 | 9.05 |
| Feature Generation | 3.71 | 2.17 | 0.08 | 20.97 | 1.21 |
| Shape Detector | 4.82 | 300.82 | 1.07 | 18.75 | 173 |
| Hough Transform for Lines | 4.52 | 40.87 | 0.09 |
41.38 | 381 |
| Connected Component Analyzer |
4.65 | 15.06 | 0.003 | 13.73 | 121 |
| Face Detection (PVP Mode) |
7.07 | 19.36 | 0.14 | 551.91 | 5100 (0.34 sec / frame) |
- Performance is measured on ADSP-BF609 processor Silicon revision 0.0.
- “Data RAM” figures include scratch, state memory but exclude memory for input and output buffers.
- The Face Detection figures are for the module Object Detector trained to detect faces.
- The memory and MIPS requirement depends on configuration parameters and input vector attributes. The values given in the table are for a typical uses case with parameters given below.
- Image width=640, height=480, thresholds are set to have percentage of edge pixels as 3%.
- Edge Detector: Edge tracing is performed in the core and all other processing is performed in the PVP.
- Hough Transform for circle: Radius range= {18- 24}, scale = 2, accumulation threshold = 150, maximum circle to detect = 3.
- Hough Transform for Lines: Theta range = {0-45}, theta step = 2, votes threshold = 80 and maximum lines to detect=10.
- Classifier: Number of trees = 250, number of classes = 15. Cycles specified in table are for one classification. MIPS calculated assuming 30 classifications per second.
- Feature Generation: ROI Width =44, ROI Height = 44, Number of vertical and horizontal blocks = 3. MIPS calculated assuming 30 ROIs per second.
- Connected Component Analyzer: Object size range = 20x20 to 100x100, Maximum number of active labels = 512, with moments calculation and maximum objects to detect = 1000.
- Object Detector configuration for face detection: Object size range = 24x24 to 100x100, scale = 1.2.
Systems Requirements
- Windows XP Professional SP3 (32-bit only).
- Windows Vista Business/Enterprise/Ultimate SP2 (32-bit only). It is recommended to install the software in a non-UAC-protected location.
- Windows 7 Professional/Enterprise/Ultimate (32 and 64-bit).
- Minimum of 2 GHz single core processor, 3.3 GHz dual core is recommended.
- Minimum of 1 GB memory (RAM), 4 GB is recommended.
- Minimum of 2 GB hard disk (HDD) space is required.
- CrossCore® Embedded Studio for Analog Devices Processors or VisualDSP++ 5.0 with the latest update.