Additional Projects

A deep dive into some areas of robotics & computer vision

CNN Eval Proj
Multi-label Multi-object Classification
Deep Learning
report / code
An evaluation study to determine simplest data and model-centric approaches for shape & colour classification in images. The objective was develop a minimalistic pipeline, engineering every last bit from pure hyperparameter tuning and training strategies, capable of robust classification.
Training strategies: Experimented the effect of Canny Edge pre-processing, CNN model depth, loss functions, data augmentation on validation loss, pipeline stability and training time.
Hyperparameter tuning: Fine-tuned fully-connected and convolutional layers to find optimal point of depth and precision.
Label-free mapping: Deployed an empirical spectral mapping for color recognition without explicit labelling. Achieved 95% accuracy on entire dataset.
Swarm Area Search using Chaos-Based Secure Communication
Bachelor's Degree Thesis
report
Designed a C++ based central system for decentralised path planning for a hetereogenous drone swarm (varying search radii and onboard capacities), with a custom chaotic systems-backed encryption network for inter-system communication.
Search radius-aware path planning: Developed a C++ based central control system, with dynamic Voronoi cell generation for search-radius aware path planning, with 57% trial accuracy in collision test-case verification.
Optimized planning algorithm: Deployed a traveling-salesman problem for modelling planning, adding HK optimization for 15% improvement in collision cases and reduced search area overlap.
Chaotic system encryption: Designed a Rossler chaotic system for two-way keyed secure modulation, with random number generated demodulation keys. Tested system against penetration and man-in-the-middle attacks to verify security.
Test drone development: Built a small (~2.5lbs) drone using Pixhawk 2.4.8 from the ground up, selected electronic speed controllers (ESCs), motor capacity, propeller radius and battery based on flight time requirements and desired range of operation.

cloned from here!