Traffic Scene Information Extraction Based On Multi-Scale Kernel Convolutional Network

Along with the rapid development of the modern cities,the Intelligent Traffic Systems(ITS),which is represented by ADAS,is becoming the principal technology to guarantee the safety and smoothness of urban traffic.The extraction of the information richly contained in the traffic scenes with cameras and intelligent computation devices is an important premise of the effective adjustment and control of the traffic flow.Currently many researchers adopt the traditional techniques of image processing and pattern recognition to analyze the image data captured from the cameras.This approach utilizes the traditional image feature extraction methods,and highly relies on the prior knowledge and professional experience to the required information,whose cost is comparably high.Besides,the extracted feature could only describe the information of certain objects from certain aspects,and is not a global way of information extraction.Aiming at solving the existing problems in the traffic scene information extraction,based on the recent research progress in the area of deep learning,this paper proposes a new network structure-multi-scale kernel convolutional network(MSKCN).This kind of network utilizes the convolution kernels of different sizes within the same convolution layer to extract the feature information from different scales.Meanwhile,along with the increase of the number of the network's layers,the abstraction of feature information is enhanced so that the description power of the feature information is improved.By adjusting the classic backpropagation algorithm,it is possible to perform supervised training of the network on the dataset,thus learning the feature representation of the objects with pertinence.The proposed MSKCN does not rely on manual design,and embraces a universality of the diverse information in the traffic scenes.By utilizing the feature visualization techniques,the learned feature can be projected into the image space,thus one can perform feature selection to eliminate the ineffective features so that the model volume could be compressed and the computational performance could be enhanced.This paper applies the proposed MSKCN to two of the most representative fields of the traffic scene information extraction.First,in the problem of road marking recognition,the proposed network can effectively recognize road markings with a high recognition rate of 97.99%,which surpasses the state-of-the-art approaches.Second,in the problem of geometric labeling of scene regions,thanks to the splendid feature extraction capability of the proposed MSKCN,the neighborhood of every pixel in the image is described so that every pixel could be geometrically labeled.By combining the superpixel segmentation result,the regions in the image are labeled.The proposed MSKCN achieves comparable result to the state-of-the-art methods with better computational performance and lower memory requirement.The effectiveness of the proposed MSKCN is validated through the two above applications in the field of traffic scene information extraction.