Cognitive Pilot

• Robust computation of visual odometry for railroad monocamera using classical CV algorithms. • Development of tramway localizer (project - ”Moscow Tramway”) taking into account the height profile obtained by spline splicing. • Asynchronous video stabilization in agricultural projects. • Development of Kalman filter for filtering static objects in local and global coordinates. • Experiment with MI metrics to compare object segmentation and its real position in the image. • Comparison of object segmentation with google map space. • Inertial sensor simulator modeling to test localizer using Kalman filter.

Huawei Technologies Co., Ltd

• Engaged in improving cloud-based game rendering in the task of information compression, extracting useful data from game engines to reduce server-to-user latency in HQ (Shenzhen, China). Used ML approaches to evaluate the impact of water effects modeled by Fresnel coefficients on the final compression quality based on MOS-learning metrics. • Made a trainable bitrate controller to solve the problem of selecting optimal geometry and texture quantization parameters at a given bitrate to achieve the best quality in a 3D data sequence compression task. Implemented a solution that gives a 10 percent improvement over the default HQ settings. • Created 3 approaches to color correction in panorama task using AI and classic Android solutions and Retinex theory, which proved successful on test data from China. • Improved the basic car environment view on TDA4 platform using OpenGL. • Made a new car environment view on Kirin 990 with neural network depth estimation on mono camera to add environment detail and remove the effect of nearby objects. The demo was successfully implemented and transferred to HQ.

HF Labs

• Developed and implemented a two-stage system for road surface segmentation in winter conditions, achieving an IoU of 97.2%, by creating an edge detection method for an autonomous road cleaner robot. • Made a tool that allows you to calibrate the external and internal parameters of the camera in 10 minutes, the accuracy of which allows you to project detections in the 3d world with an accuracy of 5 centimetres • Constructed a pipeline for data collection, annotation, training, model optimization for Jetson Orin, and production deployment, enabling daily AI model updates on the robot. • Enhanced camera performance by implementing NVIDIA GPU encoding and decoding for IP/OAK-D cameras, quadrupling the frame rate and reducing the bitrate by a factor of three. • Mentored a junior engineer who successfully trained a pole detection model from scratch to achieve a 98% F1 score.