When machines learn to see
”Detecting environments and interpreting scenarios are vital capabilities as we move towards autonomous systems. Developing computer vision therefore requires an understanding of embedded technology, algorithms and systems.”
The development and evaluation of computationally intensive high-performance image processing algorithms is often performed on powerful platforms. On many occasions, the next stage of porting them to the embedded system creates a conflict between the processing power actually available and the processing power required for implementing the image-processing algorithms. The hardware cannot cope with the software in terms of capacity and runtime. Particularly where very high volumes are involved, switching to more powerful and therefore more costly hardware is scarcely an option. Based on our expertise an analysis of applicability and the potential for improvement, we offer our customers the chance to combine performance and resources in a single cost-effective solution.
We also develop embedded prototypes at an early stage, taking account of all aspects of potential series production later on – for example regulatory compliance, unit cost, power input and output are geared towards iterative function optimization, delivering a cost-effective solution. We also rely on automated sensor data labelling and machine learning.
Where complex environments have to be detected and situations interpreted, the costs and effort for development as well as verification and validation rise significantly. Various image sources – such as RADAR, LIDAR, ultrasound and cameras – plus merging them and ensuring functional security require methods that keep costs in check and deliver efficiency. The use of environmental simulation makes it possible to create realistic conditions for verifying and validating sensor data – through ray tracing, for example – in virtual streets. Alongside the use of real data, various types of sensor can be parameterized and simulated, so that validation can be carried out early on in the development process.
Reality throws up a vast number of options and situations, and the system also has to react to them correctly. These cannot just be reproduced in simple test scenarios. Because an individual test case cannot be generated for every single scenario, any algorithms that are developed need to be particularly robust. We can develop suitable test benches to achieve the best test coverage.
- Advice on choice of sensor
- Cameras (2D, 3D)
- Ultrasound, LIDAR, RADAR
- Advice on choice of platform
- Industrial PC
- Consumer device
- Algorithm development according to A-SPICE, SPICE and Functional Security
- Porting & optimization
- Validation, incl. through environmental simulation with automatic labelling of sensor data
- Algorithm libraries
Image capture, object detection, image analysis, tracking and classification for:
- Medical devices
- Driver assistance systems
- Production facilities
Experience the virtual world
We work with you to develop a new business model, employing Virtual Reality and Augmented Reality in training, maintenance, planning as well as verification and validation. We bring future projects to life in the here and now in interactive virtual environments on various platforms such as PCs, smart phones and game consoles. User-friendliness and security take top priority for us.