What is Deep Learning?
Artificial intelligence consists of several subareas, with machine learning certainly being one of the most essential. It is about “learning” patterns or structures as well as possible by parameters of corresponding optimization algorithms in order to recognize them in further consequence. Since the 1950s, a wide variety of mathematical approaches have been developed to solve this problem. Artificial neural networks are a subfield of this development. This subfield of “machine learning” tries to imitate the properties of biological neural networks by means of mathematical functions. The basic idea of “artificial intelligence” is to break down the functions of the human brain into individual processing steps (algorithms) and to reproduce them with the help of mathematical and physical methods.
The use of increasingly deep networks can be regarded as a major innovation in Deep Learning. After early theoretical successes in the 1960s, Deep Learning is now celebrating a new boom with the availability of the necessary computing power. Due to the new application possibilities due to more computing power, the advantage of Deep Learning overtakes other Machine Learning methods.
Advantages of Deep Learning
The major advantages of Deep Learning are that the feature space is learned through training (i.e., from the data) and that Deep Neural Networks have an extremely large number of parameters. Thus, features of any complexity can be learned. In traditional machine learning, the features (the “feature space”) are predetermined by the expert and the number of parameters cannot be scaled arbitrarily. In general person recognition, for example, this means that standing persons might be recognized but sitting persons in wheelchairs might be “overlooked”. Deep Learning is particularly suitable for complex problems that usually occur in an uncontrolled environment. Using surface inspection as an example, it is far better at dealing with reflections, specularity or contamination.
Deep Learning Einsatz at EYYES
We use Deep Learning to solve complex image processing problems in the railway, automotive and industrial sectors, where environmental influences and variations can never be fully controlled. Image processing has achieved a real leap in quality as well as robustness through Deep Learning. As a result, object recognition can now be performed in real time. Cameras offer clear advantages over other sensors such as radar, lidar or similar. Our object recognition and classification technology enables new use cases for driving assistance systems and autonomous driving.
The EYYES Net is a flexible neural network that has been developed specifically for the recognition of objects in road traffic. It can be used for all kinds of applications in driving assistance. With the EYYES Net, a camera monitor system can be upgraded with “intelligence” using High Performance Deep Learning.
The EYYES Net is a pre-trained neural network that can recognize and classify the following objects:
EYYES Net Use Cases
The EYYES Net can be used for the following applications:
- Digital exterior rearview mirror with AI for proactive driver alerts in all traffic situations
- UNECE R151 BSIS: Blind Spot Information System with a minimum levels of faults positive alarms
- UNECE WP29 MOIS: Move Of Information System
- Bird perspectives
- Lane change assistants
- Detect moving objects such as pedestrians, cyclists, cars, trucks to maintain distances and intervene in driver assistance in case of danger
The EYYES Framework
With the modular EYYES framework we are able to implement prototypes for video processing systems within a short time. Included are modules for Deep Learning, in which trained models can be used. Modules can be linked and reused arbitrarily. Modules for different applications already exist and this construction kit is continuously extended.