Overview
Description
RT-RK offers cutting edge expertise on Renesas dedicated I/Ps for computer vision, deep learning and video processing. RT-RK offers consulting and services on computer vision and machine learning to optimize the performance of Tier 1 and OEM algorithm on Renesas products.
Features
- Computer Vision:
- CVe: MTMD parallelization
- IMPc: configurable pre-processing-post pipeline (100 transformations)
- PSC: pyramidal image scaling
- Memory Pipeline: DDR - scratchpad - LWM
- Deep Learning:
- CNN FE: improving configurability and usage of CNN Toolchain
- CNN FW: from Caffe model to optimized forward propagation
- Custom Layer - development: Fully Connected, Convolution 1x1, other on demand
- CNN Toolchain – customer integration support of Renesas tailored ISP for Human Vision and Machine Vision
- VisionIP:
- Stereo Block Matching IP – usage and configuration
- Optical Flow IP – usage and configuration
- Classifier IP - usage and configuration
- Image Processing:
- IMR – camera capture integration into pipeline
- ISP – usage and configuration, customization, development of advanced pipelines
- Training:
- Basic: introduction of Renesas I/Ps, tool chain, vision pipeline
- Advanced: focus on machine learning and CNN I/P
Applications
- Front Smart Camera for NCAP
- Surround View System
- Driver Monitoring System
- Autonomous Drive
- Intelligent Survey Camera
- Industrial Application
- Passenger Car, Commercial and Off-road Vehicles
Target Devices
Support
Support Communities
Knowledge Base
Videos & Training
RT-RK presents a demo of 3 ADAS algorithms (CMRS, DMS, BV) running on automotive grade target board RazorMotion by TTTech.
Algorithms demonstrated include:
CMRS - Camera Mirror Replacement System by DENSO Europe, enabling tracking of several vehicles and prediction of time-to-collision. The algorithm was optimized to run on 25 fps providing near real-time performance.
DMS - Driver Monitoring System by FotoNation, enabling tracking of driver face features and eyes, providing a direct way to prevent distractions and drowsiness, leading to a safer driving and fewer accidents. The algorithm was optimized to run on 30 fps providing near real-time performance.
Bird view algorithm by RTRK combines inputs from 4 cameras/video streams and creates an overview of the vehicle allowing the driver to monitor the surroundings and avoid collisions with nearby objects or persons. The algorithm was optimized to run on 30 fps providing near real-time performance.
All algorithms were optimized by use of powerful A57 ARM cores and Imagination PowerVR GX6650 GPU cores.
Algorithms were implemented by the use of OpenCL and OpenGL libraries.