Stop-Sign Identification Using A Convolutional Neural Network

The rapid production and societal adoption of self-driving cars generates a need for efficient and safe computer vision models. One salient example of this is the software required in self-driving cars which can identify and classify street signs. This particular image classification problem presents a unique set of success metrics. Firstly, the image classification software must have an extremely high accuracy for people to trust the efficacy of self-driving cars. Secondly, the model must generate its classification very rapidly to allow for real time reaction. And finally, the model must be widely generalizable, so the software can identify signs in unique instances that may not have appeared in the training data set. The utilization of a convolutional neural network for this classification problem offers the solution to each of these metrics. As an attempted proof of concept, I will create a convolutional neural network trained to identify whether an image contains a stop sign or not. The model will produce a binary classification indicating whether a stop sign is present in an image or not. To accomplish this, I will utilize the german traffic signs detection dataset which contains 39209 images with 43 different classes. I will also supplement this dataset with a subset of the imagenet dataset to present the network with images completely unrelated to traffic signs, as the software in self-driving cars will be confronted with every possible permutation of real world occurrence.