Technology: Machine Vision SDK
What we include in our technology
  • use of deep learning neural networks for the analysis of video frames – no tradeoff occurs between the amount/quality of information extracted and the video processing speed
  • machine-learning algorithms for classifications
  • multi-level computation framework that is able to process multiple data sources (video, sound/emotions, movement sensors)
  • self-learning mechanism - higher levels of the framework do complete the training of lower levels
  • temporal memory, modified CNN, contour analysis and time metrics as the sources for data combination for complex decision making by machines
  • detection of highlights
MACHINE VISION SDK
is a software module that allows the analysis of videos and the visual recognition of objects, backgrounds, actions and emotions in their various combinations
The module analyzes live streams and recorded videos generating time and in-video (/stream) location tags. It is able to visually recognize objects, backgrounds, actions and emotions sending them to a receiver unit or saving them as a file so a device can conclude as to what its ecosystem should do next.
Modules can be provided in two configurations - server configuration and embedded configuration:
Server configuration
It is a powerful server module – e.g. it can be deployed on a server with a powerful graphics card capable of processing 600 frames per second – for processing live streams and video files data.
Modules can be configured to run either on Linux or Windows. Minimum requirements: 16Gb of memory and 200Gb of disk space.
Embedded configuration
It is a cross-functional module for (wearable) cameras, drones and mobile phones.
One trained module with custom vision skills – particular scenes, actions, objects, etc. – requires 2-4Gb. The in-module skills are to be specified by a customer so a device will be able to use vision skills without an internet connection while other vision skills are going to be stored on the Cloud Hub as the entire skills base my be too large to fit into the available memory of a device.
Demo videos (3): real-time object recognition by Adventure Tourism and Security modules.
Modules are able to run on good budget graphics cards, can be tailored to customer-specified scenes and activities of interest
Fast algorithm
This Adventure Tourism module demonstrates the ability to detect and recognize:
  • barely visible and heavily obscured objects
  • the objects that visually merge together
Exact algorithm (2 videos)
This security module for surveillance cameras ensures high accuracy when it comes to detecting and recognizing perpetrators
Finding human contour
Source
Segmentation
Contour analysis with a capability to correct errors made by a neural network
Birds are dabbling, video with the mammals is coming
This example demonstrates 3 steps of video processing:
  • Segmentation with Neuro nets
  • Contour analysis using segmentation data
  • Approximation - creation poligonos, to optimize data transfer size