Development Tools for ADAS Design
Automotive Technology Solutions
Image processing has achieved a real leap in quality as well as robustness through Deep Learning. Thus, object recognition can be performed in real time today. Our object detection and classification technology enables new use cases for driver assistance systems and autonomous driving.
The future for development of ADAS use cases: The pre-trained EYYESNET or new development through our optimized CNN Development & Training.
EYYESNET
The EYYESNET is a neural network for visual object recognition from images, covering the main traffic-related classes (person, bicycle/motorcycle, car, truck/bus, train/tram) for automotive and railroad solutions. The network architecture has been optimized for high precision and low power consumption, especially for use on embedded platforms.
EYYESNET can be used with traditional AI inference engines or with the powerful EYYES Layer Processing Unit. In a 5-year training process, more than 5 PB of real image data was used, as well as an additional more than 10 PB of synthetic data for a wide range of weather and environmental conditions. The continuous development and training of the network for further applications, objects and environmental conditions, ensures the users of our EYYESNET the constant improvement of their applications and solutions.
Recognition of:
• Pedestrians (adults, children, …)
• Cyclists
• Wheelchair users
• Cars
• Trucks
• Trains
• Traffic light signals
SIRIUS Software Framework
With the SIRIUS modular software framework from EYYES, it is possible to implement prototypes for video processing systems in a short time. Included are modules for Deep Learning, in which the EYYESNET or other neural networks can be used. These modules can be linked and parameterized as desired. This allows a wide range of different applications of the framework to be represented.
• Detection of the environment by means of 2D-3D estimation
• Image processing pipeline for resizing, rotation, enhancement, acquisition & morphology
• CNN calculation and segmentation
CNN Development & Training
Our service offering CNN Development and Training is tailored specifically to the needs of the automotive industry. With the growing importance of artificial intelligence and machine learning in the automotive domain, our technology provides the ability to develop and train powerful and accurate Convolutional Neural Networks (CNNs).
Our approach streamlines the CNN development process to enable faster and more efficient implementation. With our toolchain, we use advanced optimization tools that allow us to fine-tune the architecture and parameters of CNN models. This leads to improved accuracy in image recognition and enables us to develop advanced driver assistance systems and other innovative functions.
Our focus is not only on the development, but also on the training of CNN models. We provide powerful algorithms and resources to train the models with a large amount of vehicle data. This significantly improves the reliability and accuracy of CNNs.
CNN optimization
With CNN optimization it is possible to reduce existing data by a factor of up to 40 without compromising the quality of the results. This service can be used individually and is not bound to a network type. The almost lossless reduction of the network makes it possible to solve complex tasks in an embedded system.
Auto Annotation Service
EYYES Auto Annotation is a service that saves up to 90% of data annotation effort during machine data entry for object recognition datasets.
A set of high quality and highly complex deep neural networks for object recognition is used to generate pseudo-labels, which are then fused and filtered by an innovative graph algorithm. The automatically annotated images only need to be manually validated. Removing image duplicates ensures that the quality of the final data set remains high, even if the data acquisition strategy is not optimized. Automatic detection of an object class not yet supported usually requires only a few thousand manually acquired images to train the highly complex detectors with transfer learning and modern “few-shot-learning” mechanisms.
