Novel Deep Learning Model for Traffic Sign Detection Using Capsule Networks

Abstract--convolutional neural networks are the most widely used deep learning algorithms for traffic signal classification till date[1] but they fail to capture pose, view, orientation of the images because of the intrinsic inability of max pooling layer.This paper proposes a novel method for Traffic sign detection using deep learning architecture called capsule networks that achieves outstanding performance on the German traffic sign dataset.Capsule network consists of capsules which are a group of neurons representing the instantiating parameters of an object like the pose and orientation[2] by using the dynamic routing and route by agreement algorithms.unlike the previous approaches of manual feature extraction,multiple deep neural networks with many parameters,our method eliminates the manual effort and provides resistance to the spatial variances.CNNs can be fooled easily using various adversary attacks[3] and capsule networks can overcome such attacks from the intruders and can offer more reliability in traffic sign detection for autonomous vehicles.Capsule network have achieved the state-of-the-art accuracy of 97.6% on German Traffic Sign Recognition Benchmark dataset (GTSRB). I. INTRODUCTION Traffic sign detection is a real world task which involves lot of constraints and complications.Even a minor misclassification of the traffic sign can lead to catastrophic consequences and can even lead to loss of life.It is implemented in various advanced driver assistance systems and in autonomous vehicles.A camera is present on the dashboard of the vehicle and it captures the real time video feed which is sampled into frames and fed to a deep learning model which is deployed inside a automotive embedded board.As the vehicle is driven in various environments,lighting conditions,speeds and geographies it is essential for the deep learning algorithm to be robust and reliable at all times.The camera can capture the traffic sign in different orientations and poses but the algorithm should be able to recognize the correct sign[4] and capsule networks are the perfect deep learning algorithm in addressing this problem. Generally Convolutional neural networks are used for all the state of the art deep learning neural network algorithms[5] in most of the image related tasks.Convolution captures the spatial information of the image using the kernel function in convolution layer. A CNN consists of input, output and hidden layers. The hidden layers further consists of convolutional, pooling, fully connected and normalization layers.