#Traffic Sign Recognition
Build a Traffic Sign Recognition Project
The goals / steps of this project are the following:
- Load the data set (see below for links to the project data set)
- Explore, summarize and visualize the data set
- Design, train and test a model architecture
- Use the model to make predictions on new images
- Analyze the softmax probabilities of the new images
- Summarize the results with a written report
** Rubric Points *** Here I will consider the rubric points individually and describe how I addressed each point in my implementation.
You're reading it! and here is a link to my project code
Data exploration:
I used the some basic output to determine 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
- The number of unique classes/labels in the data set is 43
Here is a visualization of an excerpt of the data set images:
The histogram shows that the distribution of labels in the training and validation set are very similar:
Preprocessing:
I didn't convert the images to grayscale because color is a distinctive feature for traffic signs. There are three dominant colors used in traffic signs: Blue, red and yellow, and usually only one (or no) dominant color per sign. So only the dominant color is already a feature that allows to categorize a sign into four different groups (including one for no color at all).
There were significant differences in image brightness, so I decided to normalize the images with respect to their maximum value. Normalization is generally advised to improve the performance of the optimizer. By normalizing with the maximum value, darker images are brightened up so the overall brighness becomes more homogenous throughout the image sets. This makes classification easier as brightness is not a distinctive feature for a traffic sign. Traffic signs are always as bright as possible for good visibility.
Augmentation:
The signs in the images are not necessarily aligned properly, some are rotated quite significantly. So I decided to add randomly rotated copies to the training set. For the rotation angle I chose a uniform distribution between -6 and +6 degrees. Higher rotation angles produced unrealistically strong rotations which led to a decrease in detection accuracy.
Flipping is not a good choice for augmentation, as this could even change the meaning of a sign.
Model Architecture: My final model consisted of the following layers:
| Layer | Description |
|---|---|
| Input | 32x32x3 RGB image |
| Convolution 5x5 | 1x1 stride, valid padding, outputs 28x28x8 |
| Leaky RELU | |
| Dropout | |
| Convolution 5x5 | 1x1 stride, same padding, outputs 28x28x16 |
| Leaky RELU | |
| Dropout | |
| Max pooling | 2x2 stride, outputs 14x14x16 |
| Convolution 5x5 | 1x1 stride, valid padding, outputs 10x10x32 |
| Leaky RELU | |
| Dropout | |
| Max pooling | 2x2 stride, outputs 5x5x16 |
| Flatten | |
| Fully connected | 800x240 |
| Leaky RELU | |
| Dropout | |
| Fully connected | 240x120 |
| Leaky RELU | |
| Dropout | |
| Fully connected | 120x43 |
| Leaky RELU | |
| Dropout | |
| Softmax |
I chose a learning rate of 0.001, a keep probability for dropout of 85%. The variables were initialized with the Xavier initializer. Optimization was done with the Adam Optimizer on cross entropy. A batch size of 512 and 10 epochs turned out to be an efficent choice.
I started with a very basic architecture close to the original LeNet. The first results were not satisfying and experimenting with hyperparamters didn't improve the results enough. Then I changed the image normalization to the apporach which uses the image maximum instead of 255 which improved the results. Next I added dropout layers, leaky RELUs and batch normalization which caused a decrease in performance. I removed batch normalization which caused a huge improvement. Unfortunetly the reasons for this remain unclear. The training accuracy was still significantly higher than the validation accuracy, so I decided to increase the model complexity by adding an additional convolution layer. Another optimization of the hyperparameters led to satisfying results.
My current results are:
Training accuracy: 1.000
Validation accuracy: 0.967
Test accuracy: 0.944
The web search turned out to be tedious. After I found only two proper images on the web, I took desperate measures: I WENT OUTSIDE. Outside I found lots of traffic signes which I recorded with my cell phone camera.
The images I found and the ones I recorded are all of high quality, so the detection results are good and the top softmax probability for each estimate was very high (>99% in all cases) and all estimates were correct.
This is higher than the accuracy obtained on the test set mostly for two reasons:
- The test accuracy at 94% is well in the error range of the new estimate of 100% with only six images.
- The distribution from which the new images are drawn is not comparable to the original image sets. The overall image quality of the new set is much higher.