Murtaza Safri

• Working on several courses in the field of deep learning.
• Working on migrating PointToBox deep learning lidar point cloud object detection and tracking network from PyTorch to Tensorflow 2.
• Researching innovative deep learning networks in joint RGB Camera and Lidar Point Cloud. Example: VPFNet: Improving 3D Object Detection with Virtual Point based LiDAR and Stereo Data Fusion https://arxiv.org/pdf/2111.14382.pdf
• Vision Transformers - Researching MSF3DDETR: Multi-Sensor Fusion 3D Detection Transformer for Autonomous Driving https://arxiv.org/pdf/2210.15316.pdf

• Implemented and tuned 2 Points+Normal global point cloud registration algorithm - use case in anomaly detection
• Researched and implemented changes to Fast Global Registration Algorithm to make it work with large point clouds using uniform density subsampling strategy, instead of the default grid voxelization to compute FPFH features
• Implemented changes to make Fast Global Registration speed improvements by an order of magnitude
• Did benchmark evaluations for Matrox Imaging Library's surface finder and comparison with 2 Points + Normal.

• AI and Computer Vision for Lidar)
• Plot functionality port from Labview to C++ for Doppler Lidar wind speed
• Proposal creation for Deep learning based tracking of objects in lidar point cloud data, such as using PointNet++, ShapeNet, VoxNet, Point2Box Networks for object detection and tracking
• Writing documentation for existing algorithms
• Adding support for writing NetCDF files
• Added support for Opal Lidar software for CentOS/Ubuntu and Visual Studio 2019 platform.

• Written computer vision software for robotics application, such as algorithm for detecting bins in depth images using hough transform
• Worked on software for 3d pose estimation of objects in bin, by removing 3rdparty library dependencies
• Lead development efforts in collaboration with senior researcher for the robot motion parameter optimization
• Full SDLC process followed including requirements gathering, architecture, software design, implementation and testing, and evaluation
• Wrote design document and test specification
• Converted Bayesian optimization code from Matlab to native C++ code with faster speed and comparable numerical performance as Matlab
• Researched and developed significant expertise in Bayesian Optimization methods and algorithms
• Ported state estimation algorithms in C++, such as Kalman filter
• Developed signal processing routines for Butterworth filter zero phase filtering with transient corrections in C++
• Implemented Latin Hypercube Sampling for Bayesian Optimization initialization of seed points and extended the algorithm to work with robot motion parameter optimization in the sense of finding sufficient well distributed feasible points
• Suggested improvements to machine learning algorithm design and improved speed of machine learning by using parallel processing (OpenMP)
• Improved computational complexity of hyperparameter tuning
• Designed and developed evaluation framework for testing motion parameter optimization modules
• Did comparison testing of our implementation of BO against GpyOpt and CMA-ES for synthetic functions and on actual robot.

• Written software to process data acquired from LIDAR sensors
• Researched and implemented algorithms for land water classification in LIDAR data and extended original algorithms to work with large datasets keeping mathematical and computational equivalence to original algorithm
• Depth corrected water bed points using Rodrigue's rotation formula
• Improved performance of several lidar point cloud processing algorithms with respect to execution speed and memory usage, by orders of magnitude
• Used techniques from computer vision and geometric analysis to analyze structure in point cloud data
• Applied techniques for sensor fusion to color point cloud data using RGB cameras
• Discovered hardware issues - problems with waveform digitizer by analyzing point cloud data
• Developed support for new pulsewave file format.

• Developed software to analyze magnetotellurics sensor data used for oil and gas exploration
• Wrote signal and data processing routines to process sensor data using statistical and robust methods
• Programming languages used were C/C++ for signal processing, and front end user interface using Qt Cross Platform SDK
• Managed repositories using Git and SVN, and feature request and bug tracking was done using Trac
• Perl and Python were used for prototyping and internal scripting.

Applied technique to tracking of cells in Brightfield medical images. Elastic and fluid registration and segmentation models used to develop tracking model. Link to thesis