Amund Tveit's Blog

Making Self-driving cars work requires several technologies and methods to pull in the same direction (e.g. Radar/Lidar, Camera, Control Theory and Deep Learning). The online available Self-Driving Car Nanodegree from Udacity (divided into 3 terms) is probably the best way to learn more about the topic (see [Term 1], [Term 2] and [Term 3] for more details about each term), the coolest part is that you actually can run your code on an actual self-driving car towards the end of term 3 (I am currently in the middle of term 1 – highly recommended course!).

Note: before taking this course I recommend taking Udacity’s Deep Learning Nanodegree Foundations since most (term 1) projects requires some hands-on experience with Deep Learning.

Traffic Sign Detection with Convolutional Neural Networks

This blog post is a writeup of my (non-perfect) approach for German traffic sign detection (a project in the course) with Convolutional Neural networks (in TensorFlow) – a variant of LeNet with Dropout and (the new) SELU – Self-Normalizing Neural Networks. The effect of SELU was primarily that it quickly gained classification accuracy (even in first epoch), but didn’t lead to higher accuracy than using batch-normalisation + RELU in the end. (Details at: github.com/atveit/TrafficSignClassification). Data Augmentation in particular and perhaps a deeper network could have improved the performance I believe.

For other approaches (e.g. R-CNN and cascaded deep networks) see the blog post: Deep Learning for Vehicle Detection and Recognition.

UPDATE – 2017-July-15:

If you thought Traffic Sign Detection from modern cars was an entire solved problem, think again:

Best regards,

Amund Tveit

1. Basic summary of the German Traffic Sign Data set.

I used numpy shape to calculate summary statistics of the traffic signs data set:

• The size of training set is ? 34799
• The size of the validation set is ? 4410
• The size of test set is ? 12630
• The shape of a traffic sign image is ? 32x32x3 (3 color channels, RGB)
• The number of unique classes/labels in the data set is ? 43

2. Visualization of the train, validation and test dataset.

Here is an exploratory visualization of the data set. It is a bar chart showing how the normalized distribution of data for the 43 traffic signs. The key takeaway is that the relative number of data points varies quite a bit between each class, e.g. from around 6.5% (e.g. class 1) to 0.05% (e.g. class 37), i.e. a factor of at least 12 difference (6.5% / 0.05%), this can potentially impact classification performance.

3 Design of Architecture

3.1 Preprocessing of images

Did no grayscale conversion or other conversion of train/test/validation images (they were preprocessed). For the images from the Internet they were read from using PIL and converted to RGB (from RBGA), resized to 32×32 and converted to numpy array before normalization.

All images were normalized pixels in each color channel (RGB – 3 channels with values between 0 to 255) to be between -0.5 to 0.5 by dividing by (128-value)/255. Did no data augmentation.

Here are sample images from the training set

3.2 Model Architecture

Given the relatively low resolution of Images I started with Lenet example provided in lectures, but to improve training I added Dropout (in early layers) with RELU rectifier functions. Recently read about self-normalizing rectifier function – SELU – so decided to try that instead of RELU. It gave no better end result after many epochs, but trained much faster (got > 90% in one epoch), so kept SELU in the original. For more information about SELU check out the paper Self-Normalizing Neural Networks from Johannes Kepler University in Linz, Austria.

My final model consisted of the following layers:

3.3 Training of Model

To train the model, I used an Adam optimizer with learning rate of 0.002, 20 epochs (converged fast with SELU) and batch size of 256 (ran on GTX 1070 with 8GB GPU RAM)

3.4 Approach to find solution and getting accuracy > 0.93

Adding dropout to Lenet improved test accuracy and SELU improved training speed. The originally partitioned data sets were quite unbalanced (when plotting), so reading all data, shuffling and creating training/validation/test set also helped. I thought about using Keras and fine tuning a pretrained model (e.g. inception 3), but it could be that a big model on such small images could lead to overfitting (not entirely sure about that though), and reducing input size might lead to long training time (looks like fine tuning is best when you have the same input size, but changing the output classes)

My final model results were:

• validation set accuracy of 0.976 (between 0.975-0.982)
• test set accuracy of 0.975

If an iterative approach was chosen:

• What was the first architecture that was tried and why was it chosen?

Started with Lenet and incrementally added dropout and then several SELU layers.. Also added one fully connected layer more.

• What were some problems with the initial architecture?

No, but not great results before adding dropout (to avoid overfitting)

• Which parameters were tuned? How were they adjusted and why?

Tried several combinations learning rates. Could reduce epochs after adding SELU. Used same dropout keep rate.

Since the difference between validation accuracy and test accuracy is very low the model seems to be working well. The loss is also quite low (0.02), so little to gain most likely – at least without changing the model a lot.

4 Test a Model on New Images

4.1. Choose five German traffic signs found on the web

Here are five German traffic signs that I found on the web:

In the first pick of images I didn’t check that the signs actually were among the the 43 classes the model was built for, and that was actually not the case, i.e. making it impossible to classify correctly. But got interesting results (regarding finding similar signs) for the wrongly classified ones, so replaced only 2 of them with sign images that actually was covered in the model, i.e. making it still impossible to classify 3 of them.

Here are the results of the prediction:

The model was able to correctly guess 2 of the 5 traffic signs, which gives an accuracy of 40%. For the other ones it can`t classify correctly, but the 2nd prediction for sign 3 – “adult and child on road” – is interesting since it suggests “Go straight or right” – which is quite visually similar (if you blur the innermost of each sign you will get almost the same image).