Dataset (German Traffic Sign Benchmark): https://www.kaggle.com/meowmeowmeowmeowmeow/gtsrb-german-traffic-sign
Grade: 17 out of 20
Introduction:
For some years now, several companies like Tesla, Toyota, and Mercedes-Benz, to name a few, have been researching and creating Intelligent Systems for autonomous vehicles and self-driving cars. This is possible due to Artificial Intelligence or, more specifically, to the subset of Machine Learning called Deep Learning. For the driver to let the car drive itself autonomously, the vehicle needs to understand and follow all traffic rules; thus, it should be able to interpret traffic signs and make decisions accordingly.
When trying to follow all traffic rules, the systems are involved in two main phases: the first one, to detect traffic signs along the road, and the second one, to recognize them (if some were detected). Our Deep Learning project focuses on Phase 2, traffic sign recognition (in this case, with Convolutional Neural Networks).
The image data we used to feed the models is public and comes from the “GTSRB - German Traffic Sign Recognition Benchmark” competition in Kaggle. This is a multi-class, single-image classification challenge from the International Joint Conference on Neural Networks 2011.
This data includes more than 50 000 traffic signs’ images, and more than 40 classes. In fact, the dataset is quite varying, because some of the classes have many images while others have few.
We intend to build a model mainly using the Keras library on Python, and the data referenced above, in order to classify thousands of images with traffic signs into several different categories.
We found some people claiming to have models with Test accuracies between 95 and 98%, so our main goal for this task is to achieve a Test accuracy of around 97%, or more. Also, the complexity of our model and the time it takes to run in Python will be important factors for us.
Group 2 members:
Bruno Belo, R20170735
Rui Monteiro, R20170796
Tomás Santos, R20170734
MSc: Data Science and Advanced Analytics - Nova IMS
Course: Deep Learning
2020/2021