Surya Prabhakaran

About Me

Hi, I’m Surya, M.Sc. Informatics student at the Technische Universität München, My interests lies in developing AI-driven software architectures for autonomous driving, focusing on real- time perception, decision-making, and safety optimization. My expertise includes 3D object detection, sensor fusion, and trajectory optimization using LiDAR, cameras, and deep learning. I specialize in processing large-scale sensor data for robust environmental perception, integrating SLAM, object detection, and reinforcement learning to enhance autonomous vehicle safety and efficiency. My goal is to advance safe and efficient autonomous systems by optimizing AI pipelines, deploying multimodal sensor fusion strategies, and improving active learning methodologies for real-time applications in self-driving vehicles, drones, and robotic platforms.

Projects

Active Learning with Gaussian and Heuristic Features for 3D Object Detection

July 2024 - Jan 2025

  • Developed a novel Active Learning (AL) strategy for LiDAR-based 3D object detection.
  • Designed an architecture-agnostic approach with a unified uncertainty score, combining localization and classification uncertainty.
  • Achieved an mAP of 70.34% with 25% of labeled KITTI data.
  • Containerized the entire pipeline using Docker for reproducibility and seamless deployment, enabling scalable experimentation and model training.

Multimodal 3D Object Detection for Autonomous Driving

May 2023 - Dec 2023

  • Developed a 3D Object Detection system for TUM EDGAR vehicle using camera & LiDAR.
  • Built an Autoware Perception pipeline with ROS2 & Docker, optimizing inference for low-latency deployment, critical for real-time perception in autonomous driving systems.
  • Integrated and deployed the perception pipeline on an autonomous vehicle, ensuring seamless hardware-software interaction for real-time processing.
  • Implemented CNN-based models SMOKE (camera-based) and PointPillars (LiDAR-based) models with inference times of 24.8 ms and 21.2 ms respectively.

Multi-Finger Grasping with Category-level Pose Estimations

June 2023 - July 2023

  • Predicted stable grasp points for an object in the given RGB-D image.
  • Implemented SGPA for category-level 6D pose estimation, reaching a 3D IOU of 80.1%.
  • Generated multi-scale grasp poses using DexGraspNet for the training dataset.

Inverse Rendering using NERF

Oct 2022 - Feb 2023

  • Worked on the NVDIFFRECMC (CVPR’22) architecture by NVIDIA, developing a Neural BRDF model for material optimization.
  • Enhanced denoising using a diffusion-based model, achieving a SSIM of 0.94.

Stereo Reconstruction

June 2022

  • Performed Stereo Reconstruction on the KITTI Stereo 2015 dataset using Block Matching and Semi-Global Matching for disparity map generation.
  • Utilized SIFT, SURF, and ORB for keypoint matching, triangulation, and point cloud generation.
  • Obtained an average disparity error of ∼1.3 px.

Computer Pointer Controller

  • Developed the application using Intel Distribution of the OpenVINO Toolkit.
  • This app control the movement of mouse pointer by the direction of eyes and also estimates pose of the head.
  • This app takes video as input(video file or camera) and then estimates the gaze of the user’s eyes and change the mouse pointer position accordingly.

Smart queuing System

  • Developed the application using Intel Distribution of the OpenVINO Toolkit.
  • This app detects and counts the number of people in a particular queue and redirects them to a less congested queue.
  • Worked on Intel DevCloud to implement the project on several hardwares[CPU, IGPU, VPU(NCS2),FPGA].

People Counter App

  • Developed the application using Intel Distribution of the OpenVINO Toolkit.
  • This app calculates:-
    • the number of people in the frame
    • time spent by those people in the frame
    • the total number of people counted

      and sends this data to a MQTT server.

Non-Proliferative Diabetc Retinopathy Detection

  • Performed multi-class classification for the dectection of Non-Proliferative Diabetic Retinopathy.
  • Implemented Inception Network(V1-V4,ResNetV1 and ResNetV2).
  • Achieved an accuracy of 93.78% on the test set.

Knee MRI Super-Resolution

  • Performed Super-Resolution on Knee-Magnetic Resonance Images.
  • Implemented LapSRN, SRGAN and SRCNN to perform Super-Resolution.
  • Achieved an SSIM of 0.9887 on the test set.

Experience

LiangDao

Work Student Computer Vision

July 2024 - Present

  • Developed and deployed a real-time AI pipeline for LiDAR-based 3D Object Detection and Tracking, enhancing autonomous perception for traffic management applications.
  • Integrated the system into real-world sensor networks using ROS2 & Docker on NVIDIA Jetson Orin, ensuring low-latency model inference and efficient data streaming.
  • Improved detection accuracy from 65.3 mAP to 68.1 mAP using Active Learning.
  • Implemented Software Continuous Integration (CI) using GitLab CI/CD, automating model deployment and validation for a scalable and efficient AI pipeline.

Agile Robots SE

Work Student Computer Vision

May 2023 - July 2024

  • Designed a 3D object tracking & trajectory prediction pipeline, leveraging Stereo Reconstruction & Kalman Filtering for SLAM-based localization in dynamic environments, achieving 30 FPS real-time inference.
  • Developed a real-time teleoperation system for Dexterous Object Grasping on Diana 7 and the 5-finger HIT Hand, leveraging Mediapipe & Frankmocap, achieving smooth 25 FPS performance.
  • Achieved high-speed 3D arm tracking for Diana 7 using the M3T tracker, optimizing inference time to just 12 ms, enhancing precision in robotic manipulation and automation.

Blickfeld GmbH

AI Work Student

Dec 2022 - April 2023

  • Developed and optimized a 3D Object Detection pipeline for autonomous driving using PointPillars architecture on OpenPCDet framework.
  • Deployed models using OpenVINO, reducing inference time to 50 ms per frame, enhancing real-time performance.
  • Integrated CI/CD workflows for automated model deployment, ensuring efficient updates and reproducibility in AI pipelines.

Tata Consultancy Services

System Engineer

August 2020 - March 2022

  • Worked as a full-stack developer for Citi Bank client. Developed an API for their Best Buy partner.
  • Deployed contents in Ford, Lincoln and Costco partners as a part of Pony and Costco Decommissioning projects.

Dept. of Translational Medicine and Research, SRM Medical College

Research Intern

August 2019 - August 2020

  • Collected real-life Knee MRI data, cleaned the data and used image processing techniques to make the data trainable.
  • Developed a novel loss function for the same which provided better results and submitted a paper on the same which got accepted at the EMBC 2020 conference.

SPARC, SRM Institute of Science and Technology

Research Intern

June 2019 - June 2019

  • Worked on a Research project proposed by Indian Institute of Technology Kharagpur, collaborated with University of California, Davis.
  • Developed a classifier to predict non-proliferative Diabetic Retinopathy.
  • Implemented several state-of-the-art deep learning architectures for Image Classification

Alpha Cloud Labs

Research Intern

December 2018 - December 2018

  • Worked on video text- recognition using Tesseract-OCR, OpenCV and Deep Learning.

Education

Technische Universität München

M. Sc. Informatics

April 2022 - April 2025

Thesis: Active Learning with Gaussian and Heuristic Features for 3D Object Detection (Grade: 1.0) CGPA: 1.8

SRM Institute of Science and Technology

B. Tech. Computer Science and Engineering

July 2016 - May 2020

CGPA: 8.78

Publications

Surya Prabhakaran -  - References on request