Deepak Ganesh

Automating driving tests with Computer Vision and Intelligence

Current Assignment at AI Vision through Incluso AB

Semantic Segmentation-aided Bundle Adjustment • Investigated geometry-aided semantics to enhance camera pose estimation in forest environments. • Introduced a novel cylinder model representation for localizing tree-trunks as static scene objects. • Further improved the Bundle Adjustment formulation in COLMAP by introducing an additional semantic loss. • Leveraged semantics to address challenges like dynamic objects, repetitive textures, and tree occlusions.

Visual Inertial Leg Odometry (VILO) for ICRA 2024 Quadruped Challenges • Researched and implemented robust odometry pipeline for the quadruped robodogs such as the Unitree A1 and Go1 robots. • Using existing Visual Inertial Odometry pipeline such as VINS Fusion as baseline for benchmarking as well as to tune our proposed pipeline. Calibrated and implemented the pipeline on our Unitree robot hardware with a Intel NUC computer. • Developed a fusion approach from multiple sensors and odometry information based on an Extended Kalman Filter to robustify the robot against multiple challenges such as slippage, poor exposure and lighting, etc. • Compared and evaluated the resulting pipeline against popular VILO pipeline 'Cerberus' and baseline 'VINS Fusion'

• Developed multi-object face tracking algorithms with the Tobii Glasses team. • Researched detection algorithms as proof of concept to remove sensitive personal information (primarily faces and license plates). • Implemented a robust face detection algorithm for data anonymization, based on the research work DeepPrivacy2.

Part of the perception module, working on Object tracking algorithms such as SORT and DeepSORT to track cones and substantially improve the speed of the detection pipeline.

Worked for an year at the Dynamics and Control lab on implementing a Convex optimization based algorithm for collision avoidance of a differential drive robot at Indian Institute of Technology, Madras.

Developed a mechanical deformation model for the Hard Magnetic Soft Material beams under the effect of the Magnetic field. Also derived a functional representation of the beam through a numerical approach employed using MATLAB. Presented my extended analysis of the work successfully at the end of the SURGE programme.

Robotics and Autonomous Systems track

Minor in Artificial Intelligence • Deep Learning • Reinforcement Learning • Pattern Recognition and Machine Learning • Natural Language Processing