Road Drivable Area Detection Method Based On Deep Learning

With the improvement of people’s living standards,the popularity of vehicles has increased,which raises more frequent traffic accidents.According to statistics,25% to30% of traffic accidents relate to a driver’s alertness to roads,and the merging process is one of the crucial occasions for traffic accidents.Also,the safety of driving depends on the driver’s concentration,and resulting traffic accidents are nothing new because of sharing roads between pedestrians and other non-motor vehicles on unstructured roads.All these have brought tremendous challenges to the development of the automotive industry.To solve the above problems and based on the characteristics of semantic segmentation,it has multi-scale or relational context,global to local guidance,and adaptability,this thesis proposed two road-driving area detection methods based on deep learning,MDSnet,and MAnet.In order to verify the correctness of the two models,we conducted simulation experiments.The experimental results showed that the accuracy and intersection over union(Io U)of the two methods are better than traditional models.In addition,comparing the two methods,the accuracy and Io U of the second method are higher than that of the first method,showing that MAnet model is more practical.The main research contents of this thesis are:This thesis proposed a driving area detection model of MDSnet(Multi-scale deep supervision network).Aimed at the problem of large changes in the scale of drivable areas and different scales in different scenes,features extracted by the backbone network were passed into two different extraction branches.One branch was used to take the feature of the last layer through pyramid pooling to get multi-scale features;the other branch was used to pass the information of middle layers into the global deep supervision module we designed and full used different levels of information extracted by the backbone network to get global features.In addition,to better adapt to the detection of the drivable area,the two branch modules were added with their weights,so that the model could adjust the ratio of the two parts.To verify the accuracy of the model,BDD100 K data set,and the data set collected on site were used to verify the detection ability of direct and indirect drivable areas.The experimental results showed this algorithm could more accurately detect the drivable area of roads with different scales in various weather and road conditions,and provide help for driving.In addition,we also established a drivable area detection model of MAnet(Mixed-attention network).Aiming at the problem of imbalance between drivable areas and the background samples,the model cut from another angle and used the attention mechanism to form a drivable area detection method that extracts information from the relational context.The entire network was composed of a channel domain module and a spatial domain module to strengthen the features that need to be emphasized while ignoring non-essential parts.Among them,the channel domain module used up and down sampling to extract global context information,and generated the channel domain weights together with the original features;the spatial domain module used non-local networks,and the generated features and the original features generated the spatial domain weights.Two learnable weights were adopted,which made the network self-adaptive.In addition,BDD100 K data set and domestic on-site data were used to detect direct and indirect drivable areas.The experimental results showed the model can solve the problem of sample imbalance and identify drivable areas on roads with various weather and road conditions.This thesis also designed weight balance strategies,data augmentation methods,and network parameter strategies.In addition,the corresponding modules and technologies are compared and discussed in ablation experiments.Experiments showed that the two methods proposed in this thesis could achieve good accuracy and intersection over union(IOU),and the MAnet model was better than MDSnet.Both models could solve the problems of the scale change of the drivable area in the driving scene,the imbalance of the drivable area and the background category,and the scenes of multiple weather and road conditions,which laid the foundation for the follow-up tasks of assistant and automatic driving system.This article has 56 figures,17 tables,and 89 references.