AI Development

Artificial Intelligence

Covering all steps of the process


AI-powered Applications

AI technology is already shaping various markets that we at Solectrix operate in. One example are Advanced Driver Assistance Systems.

In such applications, machine learning methods based on artificial neural networks, known as “Deep Learning”, can be used to train an AI that reliably detects pedestrians, bicycles, cars or road signs in complex traffic situations. Such an AI can be implemented as an edge device in passenger cars or as part of an intersection assistant system in trucks and buses.

Image Processing and Analysis

The World Seen Through the Eyes of an AI

A digital eye captures a traffic scene and sends the data to an intelligent control unit - the "brain". The brain processes and interprets the data, then generates a virtual image of the surroundings with classified objects. The resulting knowledge of the surroundings can be used to alert the driver to critical situations or to serve as an information source for the routing decisions of a self-driving car.


Machine Learning

How it works: Teaching the AI what to look for in an image

This level of image recognition is achieved through complex neural networks. The machine learning process used to develop such an AI involves databases with millions of tagged images, making the AI model figure out what differentiates a car from a bus, or a child from a dog. Over time, the AI model learns which image characteristics are linked to a dedicated object type, ultimately arriving at the best possible set of transfer functions for the target application’s convolutional neural network.

How it works

Technology

Implementing AI in an Embedded Project
SoC Xilinx Solectrix
Our Preferred FPGA Platforms

The heart of our AI-powered systems is usually a powerful System-on-a-Chip (SoC) by Xilinx. These two are our chips of choice:

  • Xilinx Zynq-7000 SoC
  • Xilinx Zynq UltraScale+ MPSoC

Both the Zynq-7000 and the UltraScale+ MPSoC series combine ARM CPUs with programmable logic components - they are system cores with an FPGA built-in!

Functions Implemented in FPGA
  • The FPGA component of the SoC handles the following tasks:
  • Image Signal Processing (ISP)
    This covers the processing of data from image sensors into high-quality WDR (HDR) material. Sensor fusion with other sensor technologies is also possible.
  • Safety
    Guaranteeing reliability through self-supervision routines.
  • I/O
    Handling the communication with system components, standard interfaces, high-speed memory interfaces etc.
  • AI
    An IP core is used as an accelerator for inference of Convolutional Neuronal Networks.

AI Ecosystem

The Solectrix AI Ecosystem is an original development platform for convolutional neural networks,
an advanced toolset to train a Deep Learning model. Its main fields are:
AI Ecosystem
  • Model Design
    This is where the neural network structure is defined. Usually an iterative development.
  • Data Sets
    This involves image data into the training process. Either public image databases or image data acquired by customers via their own camera hardware, labeled as ground truth annotations and separated into training data set, validation data set, and pruning data set.
  • Training Session
    Where the actual learning of network parameters is executed.
  • Supervision
    The training process is observed, creating statistical information that visualizes the development process of the training and its accuracy.
  • Model Pruning
    Where the trained network is simplified/pruned without a notable loss in accuracy while reducing the number of calculations.
  • Model Deployment
    Frozen model generation to minimize the number of operations for model inference, meaning the actual execution of the object detection in the final camera application.

Our Range of Services

Solectrix can handle or assist you with all steps of the AI development process.
  • Project Idea
    Defining the application
  • Setup Prototyping AI for Smart Data Acquisition
    Prototype implementation of a camera system for recording of training data material
  • Image Classification & Data Labeling
    Annotation of recorded material (marking of object classes and image sections)
  • Customized Model Architecture & CNN Training on Nvidia
    Setting up a training session
  • Pruning, Optimization & Quantization
    Optimizing the model to reduce number of calculations
  • Deploy Model for Inference
    The final model after training
  • Compilation to Target Platform
    Using the Xilinx workflow
  • Creation of Embedded Vision Application
    Developing the detector software
  • Final Deployment on Low-Power Devices
    Ready for the market in an embedded system