**7. Conclusion**

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 this will be to make it much more versatile and adaptable for different environments and requirements. Another benefit of increased capacity and functionality within the vehicle would be to make it less vulnerable to external access, such as hacking, which could have implications for the vehicle, its user and others.

The limitations imposed by the scale of the vehicle used in this work will affect the physical space available to house the computing hardware to that which will fit within the vehicle. This will have consequences for the functionalities the car is able to effectively demonstrate when using such small-scale computational devices.

## **Acknowledgements**

I would like to thank my supervisor Dr. Mehmet Aydin, professor of enterprise system module, for being available throughout the entire process of this project and having an in-depth knowledge of all the software techniques and development techniques required to complete this task.

I would also like to thank Dr. Larry Bull, professor of artificial intelligence, for his help in the general structure of the neural network and advice regarding the training data.

## **Notes**

This project was chosen because it is considered to be very complex and the use of neural networks to process images is a study that is always improving. Furthermore, this project was chosen as a proof of concept for autonomous vehicles, which are constantly in development for future real-world applications. This means it is an area where this theory may be beneficial for further research by others.

My intention was to improve this by scaling down the size of the neural network, making the device more portable and thus more realistic. This can be done by allowing the computations to be done on the Pi itself, making the device more mobile. It was concluded that this could potentially come at the cost of performance as trying to do computations of a large-scale neural network on a Raspberry Pi will be near impossible. As a result of this, the network needed to be scaled down, meaning much smaller images being passed, and the FPS rate will also need to be reduced to around 5–10 FPS. This is sufficient, considering the project is only a proof of concept.
