**6.3 Relation to MoSCoW and further improvements**

In relation to the MoSCoW analysis, "all of the must have", "should have" and "could have" related requirements were fully met to a good standard.

Further developments may be made as specified to cover the "would like to" aspect of the requirements.

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 generic algorithm.

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 and thus be more applicable to real-world projects.

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 to train it.

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.

**321**

*Computational Efficiency: Can Something as Small as a Raspberry Pi Complete…*

as the car rattled, wires would come free as they were not fixed in place.

and provided both stability and much greater longevity to the product.

the purpose of the current project, it was not viewed as critical.

and if it does go out of the lines, it will self-correct.

qualified under, with a score of 114/114 on the test data.

*6.4.4 Image recognition and obstacle detection*

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

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

The software design process also had problems along the way which required

The ultrasonic sensors in front of the vehicle required advanced mathematical formulae to be accurate at all distances, which would take time to complete, whereas using much a simpler formula would only mean they were accurate up to 50 cm as opposed to 5 m. This in the end was chosen as the vehicle only needs to be able to stop within 50 cm of an object, and anything beyond that distance cannot be considered an object in the vehicle's path. A useful add-on to develop the device further would be to utilise this functionality correctly for all distances. However, for

Overall, the track design was a good choice. The black background stops the edges of the paper from being seen as different to the colour of the dark floor, meaning only the edges of the track (white lines) are seen as fully qualified edges by the edge detector. The only issue with the track is that the corners can be viewed by some as "too tight" and thus the vehicle sometimes struggles to stay completely between the lines while turning through them. However, this is not a major issue as overall the vehicle remains between the lines for the majority of the driving cases,

Obstacle detection works perfectly, allowing the vehicle to stop moving if anything is within 15 cm in front of it. This is exactly what was stated in the system

Image recognition correctly classifies the images based on the direction they are

Overall, I feel that this was a successful project, in that it demonstrates a clear proof of concept that the computations required for autonomous cars do not have to be performed externally but may be done within the vehicle itself. The effect of

*DOI: http://dx.doi.org/10.5772/intechopen.88342*

**6.4 Obstacles faced**

*6.4.1 Hardware*

*6.4.2 Software*

*6.4.3 Track design*

requirements.

**7. Conclusion**

solving.

*Computational Efficiency: Can Something as Small as a Raspberry Pi Complete… DOI: http://dx.doi.org/10.5772/intechopen.88342*
