Research on Region Convolution Neural Network (R-CNN) has recently witnessed progress in both accuracy and execution efficiency in detecting objects such as faces, hands or pedestrians in photograph or video. Recently, Faster R-CNN modifies the way candidate proposals are generated in R-CNN, however, its accuracy is constrained by the size of its convolution feature map output, so it is unable to detect well those objects that are very small. In this paper, we present a fast, deep convolutional neural network based on a modified Faster R-CNN, or Visual Geometry Group 16 (VGG16) network, that we have named a Patch Proposal Network (PPN). It has a small field for recognizing objects, so that each pixel in the feature map outputs a candidate with different scales and transitions corresponding to patches. We use our PPN to detect and localize lane markings in roads. We demonstrate performance of our PPN algorithm both on the KITTI-ROAD dataset and our own traffic scene lane markings dataset. Experiments show that our algorithm is as fast as Faster R-CNN while it obtains better accuracy than Faster R-CNN in small object detection.
Pub Type: Journals
Lane Marking Detection, Intelligent Transportation Systems, Deep Learning, Neural Networks, Image Processing, Computer Vision