Development and evaluation of vision processing algorithms in multi-robotic systems.

The trend in swarm robotics research is shifting to the design of more complicated systems in which the robots have abilities to form a robotic organism. In such systems, a single robot has very limited memory and processing resources, but the complete system is rich in these resources. As vision sensors provide rich surrounding awareness and vision algorithms also requires intensive processing. Therefore, vision processing tasks are the best candidate for distributed processing in such systems. To perform distributed vision processing, a number of scenarios are considered in swarm and the robotic organism form. In the swarm form, as the robots use low bandwidth wireless communication medium, so the exchange of simple visual features should be made between robots. This is addressed in a swarm mode scenario, where novel distance vector features are exchanged within a swarm of robots to generate a precise environmental map. The generated map facilitates the robot navigation in the environment. If features require encoding with high density information, then sharing of such features is not possible using the wireless channel with limited bandwidth. So methods were devised which process such features onboard and then share the process outcome to perform vision processing in a distributed fashion. This is shown in another swarm mode scenario in which a number of optimisation stages are followed and novel image pre-processing techniques are developed which enable the robots to perform onboard object recognition, and then share the process outcome in terms of object identity and its distance from the robot, to localise the objects. In the robotic organism, the use of reliable communication medium facilitates vision processing in distributed fashion, and this is presented in two scenarios. In the first scenario, the robotic organism detect objects in the environment in distributed fashion, but to get detailed surrounding awareness, the organism needs to learn these objects. This leads to a second scenario, which presents a modular approach to object classification and recognition. This approach provides a mechanism to learn newly detected objects and also ensure faster response to object recognition. Using the modular approach, it is also demonstrated that the collective use of 4 distributed processing resources in a robotic organism can provide 5 times the performance of an individual robot module. The overall performance was comparable to an individual less flexible robot (e.g., Pioneer-3AT) with significant higher processing capability.