Road Sign Recognition

DataSet

We used a dataset from Kaggle available at this link : https://www.kaggle.com/datasets/shanmukh05/traffic-sign-cropped

In this dataset we find a total of 27 000 train images and 12 630 test images. Images are cropped and converted to .jpg format.

There is a total of 43 classes, which means that we have 43 different traffic signs:

Implementation

To speed up the training process and allow the use of GPU, we ran this project of Google Collab.

We used Pytorch to design our model and our datasets class

We tried different model trying different number of hidden layers, different hidden layers's width and different convolution net.

The following results were obtain with 20 fully connected hidden layers of size 150. We added as activation, the Tahn function at each node, and used the dropout feature with a value of 0.25 Here's the detail of the convolotion net and the neural net:

SignClassification(
  (c1): Sequential(
    (0): Conv2d(3, 5, kernel_size=(2, 2), stride=(1, 1), padding=(1, 1))
    (1): BatchNorm2d(5, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
    (3): AvgPool2d(kernel_size=2, stride=1, padding=1)
  )
  (c3): Sequential(
    (0): Conv2d(5, 10, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (1): BatchNorm2d(10, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
    (3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (c5): Sequential(
    (0): Conv2d(10, 28, kernel_size=(2, 2), stride=(1, 1))
    (1): BatchNorm2d(28, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (c7): Sequential(
    (0): Conv2d(4, 7, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (1): BatchNorm2d(7, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
    (3): MaxPool2d(kernel_size=2, stride=1, padding=0, dilation=1, ceil_mode=False)
  )
  (layers): Sequential(
    (0): Linear(in_features=1372, out_features=60, bias=True)
    (1): ReLU()
    (2): Linear(in_features=60, out_features=60, bias=True)
    (3): ReLU()
    (4): Linear(in_features=60, out_features=60, bias=True)
    (5): ReLU()
    (6): Linear(in_features=60, out_features=60, bias=True)
    (7): ReLU()
    (8): Dropout(p=0.3, inplace=False)
    (9): Linear(in_features=60, out_features=60, bias=True)
    (10): ReLU()
    (11): Linear(in_features=60, out_features=43, bias=True)
  )
)

We used Cross entropy as Criterion and the Adam optimizer for faster and accurate training.

Results on test set

Epoch [1/8], Step [1/205], Loss: 3.7597
Epoch [1/8], Step [50/205], Loss: 3.0578
Epoch [1/8], Step [99/205], Loss: 1.9029
Epoch [1/8], Step [148/205], Loss: 1.9718
Epoch [1/8], Step [197/205], Loss: 1.2675
Epoch [2/8], Step [1/205], Loss: 1.2280
Epoch [2/8], Step [50/205], Loss: 0.9330
Epoch [2/8], Step [99/205], Loss: 0.9221
Epoch [2/8], Step [148/205], Loss: 0.6355
Epoch [2/8], Step [197/205], Loss: 0.5433
Epoch [3/8], Step [1/205], Loss: 0.3476
Epoch [3/8], Step [50/205], Loss: 0.3307
Epoch [3/8], Step [99/205], Loss: 0.2422
Epoch [3/8], Step [148/205], Loss: 0.2220
Epoch [3/8], Step [197/205], Loss: 0.2694
Epoch [4/8], Step [1/205], Loss: 0.1245
Epoch [4/8], Step [50/205], Loss: 0.1606
Epoch [4/8], Step [99/205], Loss: 0.1115
Epoch [4/8], Step [148/205], Loss: 0.1777
Epoch [4/8], Step [197/205], Loss: 0.2607
Epoch [5/8], Step [1/205], Loss: 0.2845
Epoch [5/8], Step [50/205], Loss: 0.1755
Epoch [5/8], Step [99/205], Loss: 0.1066
Epoch [5/8], Step [148/205], Loss: 0.1257
Epoch [5/8], Step [197/205], Loss: 0.1497
Epoch [6/8], Step [1/205], Loss: 0.1188
Epoch [6/8], Step [50/205], Loss: 0.0705
Epoch [6/8], Step [99/205], Loss: 0.0656
Epoch [6/8], Step [148/205], Loss: 0.0850
Epoch [6/8], Step [197/205], Loss: 0.0337
Epoch [7/8], Step [1/205], Loss: 0.0345
Epoch [7/8], Step [50/205], Loss: 0.0622
Epoch [7/8], Step [99/205], Loss: 0.0201
Epoch [7/8], Step [148/205], Loss: 0.0400
Epoch [7/8], Step [197/205], Loss: 0.0467
Epoch [8/8], Step [1/205], Loss: 0.0356
Epoch [8/8], Step [50/205], Loss: 0.0281
Epoch [8/8], Step [99/205], Loss: 0.0203
Epoch [8/8], Step [148/205], Loss: 0.0169
Epoch [8/8], Step [197/205], Loss: 0.0114
Finished Training
Accuracy of the network on the 12630 test images: 89 %

We found a pretty high accuracy of our model on the test set

Then we tried to figure out, wich signs were wrongly classified:

We provide an heatmap highlighting which signs are wrongly classified and what's the predicted class for these

Some of the highest value outside of the diagonal are class 41 classified as class 9 for 1/3 of the observations, class 0 classified as class 4 for 1/3 of the observations and class 27 classified as class 18 on 30% of the test set.

Class 41: Classified as class 9: