Traffic Sign classification requires a high degree of accuracy such that it is viable to use in autonomous vehicles.
we try to explore the different features of traffic sign images and use state of the art machine learning algorithms to classify traffic signs for autonomous cars. Our model is a variation of VGG-16 which uses depth of the image to accurately classify image.
Final Report can be found here.
Lisa dataset is a collection of annotated im- ages and videos containing traffic signs and videos containing traffic signs. It is comprised of over 6,000 frames that con- tain over 7,000 signs of 47 different types.
The model is based on VGGNet. Overall High level architecture of our model is given below,
We have used VGGNet pre- trained model to speed up training process
| Data Set | Number of Classes | Accuracy | Testing time | Train Time(per epoch) |
|---|---|---|---|---|
| GTSRB | 43 | 99.3% | 15.3 ms | 210 sec |
| GTSRB(GPU) | 43 | 99.31% | 3.5 ms | 45 sec |
| LISA | 16 | 98.7% | 14.6 ms | 33 sec |
| LISA(GPU) | 16 | 98.76% | 3.5 ms | 3 sec |
We see a considerable difference between the training and testing time of CPU and GPU Convolutional Neural Net- work. With the German Traffic Sign dataset we got an accuracy of 99.31% and with LISA dataset 98.7%. These differences are due to the imaging conditions like weather, speed of the vehicle, occlusions and light conditions.
Some of the images are not clear and to improve our model , we plan to integrate object segmentation with classification to accurately identify traffic signs on real-time.
[1] Simonyan, Karen, and Andrew Zisserman. ”Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556 (2014).
[2] P Sermanet , Y LeCun . Traffic Sign Recognition with Multi-Scale Convolutional Networks .
[3] VGGNet - VggNet Github (https://github.com/machrisaa/tensorflow-vgg)