Technologies: Python, OpenCV and others Adaptation of advanced segmentation and detection models for accurate recognition of furniture and other objects in images. This has significantly improved the accuracy and quality of object selection. Modern contour post-processing technologies have been applied, which have improved the detail and ensured the exact overlap of elements such as wallpaper, floors and ceilings. The use of the diffusion model made it possible to completely remove objects from the images, restoring in detail the missing voids of the room.

Technologies: C++, OpenCV   The goal of the project is developing software for fast automatic calculation number pixels located in color limits. The video system captures 30 cm x 30 cm images of falling objects - at a speed of 5-to-50 km/hr speed with a Logitech C920 camera. The aim is to calculate how many pixels fall within a band of color [e.g. RGB start (255,0,0) to (200,50,50) i.e. R will range from 200-to-255, G from 0-to-50, and B from 0-to-50]. The developed program returns the count of pixels for each image (frame) at a given constant rate - e.g. 10 frames/second-to-60 frames per second.

Technologies: OpenCV, Yolo2/Faster RCNN / Mask R-CNN, COCO, Jetson Xavier   The aim of the project was to implement traffic counting (multiple object detection + tracking)  for installations placed in the countryside with low power consumption requirements. Using background subtraction, deep neural networks and other methods we optimized the models to run on Jetson Xavier hardware platform meeting the clients’ requirements. Significant part of the job was to adapt the computation environment to the hardware.

Technologies: Python, OpenCV Duration: 6 months The goal of the project is to analyze the video of a tennis game for breaking match into shorter videos: one video per point. It was required to remove those parts of the match where the players did not play (the players rest, the gap between the points, etc.); that allowed game statisticians to make further revisions of the game much faster because all "idle" periods of the game were removed and the total length (as soon as file size) was much shorter. The logic of breaking video has been developed based on the analysis of game events that were detected in the video, position, speed and posture of players, ball movement and location, and other parameters. CV algorithms were used: optic flow, background subtraction, HoG detector, pose detection and others.