Lane And Vehicle Detection Based On Deep Learning

With the development of the times,automatic driving has attracted wide attention recently.The detection of lane and vehicle is one of the most important tasks in automatic driving.Due to the complex and changeable driving environment,traditional statistical learning methods can no longer satisfy the requirements of the new era of automatic driving.It is imperative and extremely challenging to study high-performance detection algorithms.Convolutional neural network is a feature extraction method that can effectively characterize the semantic features of regions,providing a new idea for vehicle and lane detection.Based on deep learning technology,the major work of this paper is summarized as follows:The object detection networks based on convolutional neural network are studied.By building the vehicle training data set and the detection networks,the models of vehicle detection are generated.The models are tested on the difficulty and scale levels,which proves that the model can effectively improve the vehicle detection performance compared with the traditional statistical learning methods.Meanwhile it provides a reference for the subsequent optimization network.In order to overcome the difficulty of detecting small sized vehicle targets often found in the detection model,this paper introduces the method of object-scale analysis and obtains distribution characteristics of the object’s scale in the KITTI and PASCAL VOC data sets.The region proposal network is optimized by using the distribution characteristics.At the same time,a multi-scale convolution kernel is added to the model to improve the multi-scale adaptability of this model,and a residual network structure is applied to optimize the gradient transfer.All of the above strategies can effectively improve the detection performance of the model.The lane detection algorithm based on threshold segmentation and curve fitting is first studied and implemented.An annotation algorithm is designed to build a fine-labeled lane data set for lacking of lane data set.Based on the full convolutional network and the conditional random field graph model,the segmentation model is built and trained for lane detection,which can segment lane on pixel level.Finally,a fusion model of lane and vehicle detection network is studied.Specifically,first,features are extracted by sharing convolution features.Then the lane and vehicle detection networks are added behind deep feature map to realize detection of lane and vehicle simultaneously in one model.the model can simplify the model and increase the detection speed obviously while completing detection task.