Arko Sarkar

AI + SoC

Optimizing computer vision models for the edge.

- Designed End-to-End solution for Bird's eye view (BEV) Perception algorithm both training and deployment on Nvidia Orin AGX.

- Selection of algorithm and trained BEV segmentation(free space and boundary) on custom dataset of mines.

- Architected everything for inference on Nvidia Drive Orin SoC.

- Optimization of overall pipeline.

- Inference hardware and camera selection.

- Camera Calibration

- Data Collection strategy support.

Pytorch, C++, TensorRT C++, OpenCV cuda, cuda kernel, Driveworks, DriveOS, Computer vision

• Multi object tracking using DeepSORT.
• [DSP Programming OpenCL] Development of computer vision, image processing and robotics post-processing kernel on Synopsys EV74 Vector DSP( Computer Vision IP).

e.g. Implemented highly parallelized computer vision function from scratch on Vector processor(DSP) like WarpAffine, crop_resize, nms, cosine_d, cv pre/post-processing algorithms(yolo v5/v7/x, centernet, reid, ggcnn, segmentation, inpainting etc). Achieved highest inference runtime for the aforementioned functions.

• Computer vision inference
• Optimise computer vision pipelines
• OpenCV

• Project 1 - Localization (Level 3 Automated Driving)
• Technically leading a team of 9 engineers.
• Algorithm development of Localization (Function) component using High Definition Maps (HERE Maps). Algorithm based on
  statistical signal processing, state space estimation and probabilistic models.
• Product development from PoC to production software. Software development, validation, verification and in- vehicle testing.
• Concept and development of an in-house simulation software to evaluate localization in all possible corner use cases. GUI based tool
  that can model complex road scenarios and sensor (Camera, Lidar, Radars, GPS) data using simple sensor model. Responsible for
  product requirement, software requirement and algorithm selection. End to end development with external supplier.
  
• Project 2 - Functional Evaluation (Level 4 Autonomous Driving)
• Building and scaling an evaluation framework infrastructure to evaluate deep learning based perception and sensor fusion
  algorithms. Development of evaluation metrics and strategies for deep learning models. Benchmarking against state-of-art
  evaluation of computer vision community.
• Technically Leading a team of 11 engineers.
  Hackathon conducted across India for solving the problem of "Generation of Lidar point cloud data".
• Recruitment of post graduate engineers for Department.

• Development of platform software for heads-up display.
• Renesas V850-DK4H Graphic Library enhancement. Understanding of graphics driver and graphics hardware components.
• Development of service layer for stepper motor complex driver development.
• Power Management of ECU.
• Development of safety code for gear display.
• Involved in development of various application software modules for Heads-Up-Display.
• Travelling across Visteon (Paris, Yokohama- Japan) and Honda R&D (Utsunomiya Japan) to conceptualize product requirement and
  demonstrate iterations of software to Honda-Japan.

• Implementation of Graphical Framework development (MVC Architecture) for instrument cluster.
• GUI screen development using for various features of Instrument cluster.
• Athlet ImageM graphics library.
• Multiple travel to Sofia for onsite integration.