Software architecture and computer vision for resource constrained robotics

This thesis identifies the restrictions that resource-constrained robotic platforms experience in relation to their software architecture and computer vision. It proposes and evaluates a number of techniques to support more effective implementations. A review of robotic software shows that they often implement a message passing framework as their software architecture. This supports maintainability and modularity. However, the loose coupling afforded by message passing systems has a computational cost on performance and inhibits the ease of accessing multiple data sources. Resource-constrained robotic systems do not have the computational power available to support real-time application of a message passing framework. To resolve this, two compile time message routing techniques, Typemap Routing and Link Time Route Generation (LTRG), are presented and evaluated. They are compared against a hashmap and treemap in both single and multithreaded environments and were shown to improve predictability and latency. A technique called co-messaging is identified to reduce complexity in message passing systems through easier data access. Using Typemap Routing and co-messaging, NUClear, a software architecture framework was developed. NUClear's execution time was compared to Robot Operating System (ROS) and through an implementation of NUClear in the NUbots' codebase. NUClear's interface complexity and memory usage were evaluated, showing improvements. The implementation of co-messaging and Typemap Routing supported the development of an Adaptive Lookup Table. It was developed to resolve the issues that resource-constrained robotic systems experience when classifying colours in various lighting conditions. It was assessed against and outperformed a static lookup table. A Visual Mesh was developed to improve computer vision and reduce the computational cost of convolutional neural networks on resource-constrained robots. It was evaluated against a hexagonal mesh and other convolutional neural networks and resulted in improved accuracy and execution performance. Its performance over distance exceeded more complex networks. The techniques developed and the evaluation presented in this thesis are important for the fields of software architecture and computer vision. They offer solutions to improve performance for both resource constrained robotic systems as well as less constrained systems.