Research On Location Of Road Defects In One Network Architecture

With the rapid development of road network,the road safety has attracted increasingly attention.There will be a lot of damage when the road is rolled by vehicles,damaged by weather and not effectively maintained for a long time.Road defects not only affect the walking of pedestrians,but also affect the driving of vehicles.The traditional manual detection of road can not find faults in time.At the same time,it also consumes labor,and the accuracy can not be guaranteed.Before the development of machine learning and computer vision,the road defect detection algorithm based on digital image processing has been widely used,but it still can not solve the problem of accuracy.Therefore,how to detect road defects quickly and accurately has become a hot issue.In view of the above problems,the main work contents and innovations of this paper are as follows:In the work of this paper,firstly,mask RCNN,cascade RCNN and other instance segmentation algorithms are used to segment road defects,focal loss is used to reduce the problem of sample imbalance,and multi-layer feature pyramid is introduced to obtain more detailed road defects features.Secondly,lanenet and other lane detection algorithms are used to detect the lane of the road.Finally,through lane information and defect instance segmentation information,the specific lane in which the road defect is located can be quickly located.In order to improve the forward inference speed of the model and achieve multiple tasks in a network,we can detect the lane on the road while segmenting the road defect in a network,so as to determine which lane the road defect is in.This paper proposes a new multitask training method and a new model.The model includes a shared feature extraction network and two detection heads(lane detection head and road defect instance segmentation head).By alternately training different network heads and continuously reducing the learning rate and epochs,the network model proposed in this paper can achieve the two tasks of instance segmentation of road defect and lane detection,so as to determine which lane the disease is in.In the experiment of this paper,the training method is compared with the other two training strategies.At the same time,the training results of the network model are compared with the training results obtained through the single task model.It is found that the network model proposed in this paper can obtain high accuracy,reduce the internal storage of the model and accelerate the speed of inference.In addition,this paper also conducted an exploratory experiment to achieve two detection subtasks in a network by means of transfer learning and alternating training when there are only ready-made data sets and lack of annotation.