*6.4.1 Hardware*

*Control Theory in Engineering*

the first four out of five aims.

challenging and thus was not achieved in time.

aspect of the requirements.

a generic algorithm.

**6.3 Relation to MoSCoW and further improvements**

and thus be more applicable to real-world projects.

moving.

Can the vehicle utilise collision avoidance via ultrasonic sensors?

Can the vehicle recognise and stop at stop signs for a certain amount of time? Following the test phase, the vehicle has shown the capabilities required to pass

The vehicle fully utilises the neural network to accurately classify the real-time image to decide upon and then implement a movement instruction. This allows the vehicle to follow the track given any set layout and follow it until the procedure is manually stopped. A variety of obstacles have been placed in front of the vehicle during operation, and when this occurs, the ultrasonic sensors recognise it, and the vehicle stops within the specified distance in order to avoid collision. The vehicle remains stationary until the obstacle moves away or is removed from its path. If the obstacle is moved around incrementally, it will "follow" the object staying at a minimum of the specified distance away. This is a very useful functionality for potential additions to the project which is to have the vehicle following other moving vehicles or staying a specified distance away from the obstacle should it be

However, the stop sign detection capability was not implemented due to time constraints. To do this a Haar cascade qualifier must be used; this is an entirely different machine learning algorithm and thus would have to be implemented alongside the neural network. This naturally would have added many further functionalities. This was placed into the "would like to have this requirement later to develop the system design further" section of the MoSCoW analysis of the requirements because not only would this further implementation be time-consuming but finding small-scale stop signs that would be adequate for this task proved to be

In relation to the MoSCoW analysis, "all of the must have", "should have" and

Further developments may be made as specified to cover the "would like to"

Implementing a backpropagation algorithm manually as opposed to relying on the one provided from OpenCV, this would be desirable as it has the potential to be developed in a manner that may be more efficient for the problem at hand than just

A possible improvement would also be integrating the capability of designing, recognising and acting upon stop signs via a Haar cascade qualifier. This would be a further development to the device which could allow it to have more functionality

Another adaptation which would have been interesting to implement would be using a K-nearest neighbour algorithm for the machine learning part of the project. This could be used to compare the results from this to that of the neural network, to see which is more accurate. However, as aforementioned this was not implemented currently because of the much larger data set that would be required

The only final improvement which could be made to the system would be to improve the mathematical functions applied to the ultrasonic distance sensors, which will be discussed below. These are currently only accurate at reading distances up to 50 cm using a simple mathematic function which for the present problem is acceptable, but future adaptations to the device may require more accuracy

and thus more advanced mathematics to allow it to be accurate up to 5 m.

"could have" related requirements were fully met to a good standard.

**320**

to train it.

Hardware design encountered a variety of problems with the circuitry of the vehicle which connected the external components to the Pi. Using a breadboard to connect all the components was not viable in the long term, and testing proved that as the car rattled, wires would come free as they were not fixed in place.

This was fixed by having a permanent version of the circuit made and soldered in the workshops at the university, producing a finished product seen in **Figure 7**. This took a while to be perfect, and because of this, other tasks had to be put on hold pending the new board design. The new board completely solved the problem and provided both stability and much greater longevity to the product.
