Image recognition

System for comparatively analyzing and calculating Siberian tiger populations

Project targets

Develop a control system for Siberian tiger populations and movement with the help of computer vision that identifies specimens using camera traps.

The job

The team built a system for automatically identifying tigers using a specific and unique number or name in the images it gets from the camera traps. The system then lets users input information for each tiger: the unique ID, name, gender, age, number of times they’ve been photographed, the places they’ve been photographed, links to family members, ability to draw familial ties. Tigers are identified using computer vision algorithms built on convolutional neural networks.

Object detection and facial comparison systems for the police

Project targets

Develop a flexible system for recognizing a variety of objects, automating image analysis for the police. The system supports the addition of new or expanded object classes.

The job

We built a system that detects objects that match previously uploaded data sets for different object classes. A convolutional neural network drives detection, while users can perform criminal and suspect searches, add pictures, and delete pictures using the facial comparison system. In addition, there is a tool for adding new object classes.

Retinal recognition

Project targets

Develop technology for recognizing the human iris.

The job

We built algorithms for the entire biometric process used for iris recognition, which takes place over three stages:
  • segment the iris and highlight textures
  • highlight singularities and create biometric patterns
  • combine the patterns

Segment objects on communication towers in a 3D cloud of points

Project targets

Build a process for segmenting objects (well-known antenna types and models) located on communication towers.

The job

After studying the problem, we created algorithms that segment objects into a 3D cloud of points. The solution works through a series of steps: processing the cloud of points using inputted data (the entire cloud of points that makes up the tower surrounded by land), pulling the tower out of the cloud, extracting the tower platforms, and then processing each platform’s cloud of points independently. Once that is complete, a set of algorithms kicks off to process each cloud in order to segment the objects on each platform with a high degree of accuracy.