Faster R-CNN for small traffic sign detection

Traffic sign detection is essential in autonomous driving. It is challenging especially when large proportion of instance to be detected are in small size. Directly applying state-of-the-art object detection algorithm Faster R-CNN for small traffic sign detection renders unsatisfactory detection rate, while a higher accuracy will be performed if the input images are upsampled. In this paper, we first investigate Faster R-CNN’s network architecture, and regard its weak performance on small instances as improper receptive field. Then we augment its architecture according to receptive field with a higher accuracy achieved and no obvious incremental computational cost. Experiments are conducted to validate the effectiveness of proposed method and give an comparison to the state-of-the-art detection algorithms on both accuracy and computational cost. The experimental results demonstrate an improved detection accuracy and an competitive computing speed of the proposed method.