Visual perception is the foundation of vision-based inference. Thus, it represents the key enabling technology for mobile robotics, outdoor navigation, autonomous driving and many other contemporary CPS that are finding their way in digitized agriculture. Embedded vision platforms are the next step in the design and development of devices that provide such capabilities. They offer seamless integration in existing platforms and are highly optimized for the execution of computer vision algorithms. Small dimensions and efficient use of resources, at low cost, are always contrasting performance and diversity of capabilities that device offers. In that sense, it is hard to find a platform that would be a one-size-fits-all solution. General vision tasks ’in the wild’ are always considered as hard and the AI semantic analysis of 3D scenes has been a challenging and unanswered issue for many years. ViTech technology is capable of providing uncertainty reduction by implicit or explicit use of 3D information and is expected to bring complex vision functionalities to agricultural sensor systems operating in uncontrolled environments. It will enable the product positioning among different solutions competing on the market. Embedded platforms capable of providing necessary level of 3D perception with low energy consumption will be a preferred choice for solving vision tasks in all future CPS. The project raised the level of the technology from TLR 4 to TLR 7.

The ViTech PAE has made significant breakthrough in design of generic and modular computer vision gateway, capable of adapting to different user need and application scenarios. It is a fully customizable tool for design of real-time adaptive processing pipelines, according to user needs and with minimal programming effort. Also, it is a software package tested on commercially available hardware, which was shown to be easy for integration with commercially available drone platforms. In particular, developed system and test results provide the technology receiver with following capabilities:Full control over configuration of each of the functional blocks reflecting original system design requirements Ability to easily extend different parts of the system, due to modular design and use of open source software Ability to easily integrate different in-house solutions or external, proprietary systems and sensor devices into real-time processing pipelines created by developed vision gatewayOvercome the need of using complex, proprietary stream processing libraries, in order to achieve the same level of functionalitySupport different types of input sources, as well as different types of input stream formats in unified and systematic way (e.g. by utilizing local input streams coming from camera mounted on drone, agricultural machinery, or vehicle; or remote signals streamed to ViTech gateway over the network) Capability to exploit hardware encoding and decoding on the ViTech gateway for any combination of user defined multiple input and multiple output streams Capability to perform resource efficient deployment of AI models in outdoor environments, by utilizing heterogeneous computing of ViTech gateway and provided set of model translators and compilers enabling creation of optimized AI inference engines. Stream processing and AI deployment with low power consumption and low memory footprint Versatile tool for performing both low-level and high-level vision tasks Real-time monocular depth perception on drone platform