We are a team of computer vision and machine learning guys. We make software and hardware solutions for smart city and intellectual transport systems. In a search to join new ventures where we can use our experience, e.g. retail, home security, public access control, manufacturing as well as smart traffic and self-driving cars.
📧 [email protected]
CTO/Team lead, 1 PM, 1 data science engineer, 2 ML developers, 1 image labeling guy, 1 video processing expert, 2 senior back-office developers, 1 junior back-office developer, 1 CG developer (synthesized), 2 senior hardware engineers (embedded systems), 1 MLOps/DevOps.
We use Python with C++ for critical and performance demanding parts; TensorFlow and Torch frameworks with performance optimization for Nvidia CUDA using TRT (for embedded devices as well), Intel NCS with OpenVINO (assume outdated), TFLite for mobile devices. For object tracking we use a custom-made multichannel tracker that helps to follow an object, with some mods can also be used to track precisely other types of objects (e.g. pigs in a barn for health control). We had some math implemented for back and forth transformations into real world and camera plane coordinates and for auto-calibrate camera with limited real-world data. Transport objects detector is based on yolo4-csp scaled networks, objects classification Conv2d + MaxPool2D, ANPR recognition via custom architecture based on RCNN with ctc beam search string generation.
It’s a serverless FullHD/4К camera with Nvidia Jetson NX on board for neural and back-office fines ‘paperwork’ processing. We use precise NNs for object recognition (vehicles and people) in a video stream, then combine them into real world objects and map it into real world space. Hooked to the vehicle’s edges and wheels we calculate its speed, and also detect lane control and other types of traffic violations. Camera reads plates with a custom ANPR module, recognizes vehicle model and make (beta), performs auto-calibration when installed and periodically to compensate tilt and other hardware drifts. Our traffic complex has been government certified as a speed measuring device with +/- 1 km accuracy.
It’s a 3 years span project from scratch to operation of dozens of cameras in multiple countries with different weather conditions. With an extensive online video processing background we did a major switch, so, over the time, we've built some tools and utilities for ML/CV projects: data labeling, Unity™ based synthetic image data generator, multiple NNs architectures developed and modified, camera control tools, abnormalities detectors and compensation algorithms integrated. On the business part we did an enormous job to integrate the camera into external administrative and police databases, smart city traffic control systems and weight control stations.