Road Remnants Detection Based On Deep Learning

Road remnants are obstacles that exist on the road and affect road traffic.With the acceleration of our country’s urbanization process,the total number of road mileage in our country and the number of citizens’ car ownership continue to rise,and the hidden traffic hazards and safety threats caused by road remnants have begun to receive attention.Because of the sudden formation of road remnants,various types and few commonalities in features,it has been difficult to make progress on the automatic detection of road remnants.Aiming at the characteristics of road remnants,this paper proposes a road remnant detection method based on deep learning multi-step combination.This method relies on the image acquisition equipment of road inspection vehicles to detect road remnants,which can greatly improve the detection efficiency of remnants.According to the prior knowledge of road remnants,road remnants can be defined as abnormal objects on the road surface that are non-vehicles,non-pedestrians and can cause hidden traffic hazards.Therefore the basic idea of the road remnants detection method is proposed: The first step is to remove the background of the image and extract the road section to be detected as the region of interest.The second step is to extract the characteristics of typical non-legacy objects from the region of interest and then detect and remove them.The third step is to perform an equal size of the image Grid division.The grid is designed as a100*100 rectangular frame.After the road grid and the non-road grid are classified,the detected non-road grid can be considered as a part of the legacy.After the grid is clustered,the remnants can be obtained.The first step is to eliminate the background of the road.This paper designs a background elimination network based on the neural network resnet34.The network is composed of a feature extraction network combined with a context aggregation module.The road feature map is extracted and the feature map is enhanced to output the background elimination result.After experimental verification,the network can better segment the road segments and background that need to be detected.Compared with the traditional image segmentation network,the average intersection score is increased by 2.44%,and the result of the enhanced representation is compared with the average intersection of the coarse segmentation results.The score increased by 3.04%.The second step is to remove the typical non-legacy objects in the road.This paper proposes a typical non-legacy object removal method for road legacy detection,and designs a U-shaped backbone network that can obtain multi-scale feature maps and its loss function.It is verified that the network improves the detection speed while maintaining the detection accuracy,especially for the detection of small targets,which is 2.03% higher than the traditional detection model.The third step is to achieve the detection task of candidate remnants.This paper proposes a method for screening candidate remnants based on support vector machines.Non-road grids are obtained through grid division and classification,and connected non-road grids are clustered as the output of the detection frame of the road remnants.Experimentally verified,in the self-photographed experimental data test set,the road remnant detection method based on deep learning has a detection success rate of 92.89% for five different road remnants.This paper comprehensively discusses the road remnant detection method based on deep learning from the aspects of detection method design,network design and experimental analysis.The experimental results in the road remnants image data set confirm that the method used in this paper has high precision and strong generalization ability for road remnants detection,and has certain theoretical and application value for the research of road remnants detection methods today.