ECE 539 Project Road Traffic Sign Recognition Bo

ECE 539 Project Road Traffic Sign Recognition Bo Peng bo. peng@wisc. edu

Background & Motivation • Traffic signs ensure the safety when people drive • Automatic recognition for autonomous driving (Data. Camp)

Data • Belgian Traffic Signs Dataset (64 categories ) Training: 4575 images / Testing: 2520 images

Data • German Traffic Signs Dataset (43 categories ) Training: 39, 209 images / Testing: 12, 630 images

Method • Data transformation ØResize all the images into the same size, e. g. , 64 × 64 ØNormalize the greyscale of the images to [0, 1] • Divide training data into training part and validation part ØSize of validation data: 20% • Convolutional neural network + fully connected layers

Results and Discussion • Training process • Epoch: 1. . . Train acc: 50. 0%. . . Validation acc: 62. 5%. . . Validation loss: 1. 245 • Epoch: 2. . . Train acc: 90. 6%. . . Validation loss: 0. 245 • Epoch: 3. . . Train acc: 96. 9%. . . Validation loss: 0. 117 • Epoch: 4. . . Train acc: 100. 0%. . . Validation acc: 96. 9%. . . Validation loss: 0. 183 • Epoch: 5. . . Train acc: 100. 0%. . . Validation loss: 0. 029 • Epoch: 6. . . Train acc: 96. 9%. . . Validation acc: 90. 6%. . . Validation loss: 0. 179 Acc = percentage of images being correctly recognized Loss = softmax cross entropy between truth and predictions Epoch 6 is over-fitting

Results and Discussion • Belgian Traffic Signs Image size Num of epochs Batch size Testing acc 64 7 16 89. 7% 64 7 32 86. 0% 64 79. 8% • German Traffic Signs Image size Num of epochs Batch size Testing acc 32 6 16 89. 7% 32 6 32 87. 0% 32 5 32 88. 0%