Research On Nighttime Road Obstacle Image Semantic Segmentation Based On Attention Mechanism

Obstacle detection is very important for the safety of driverless vehicles on the road.Based on the existing problems in obstacle image segmentation technology,such as low segmentation accuracy or low segmentation speed,and insufficient application scenarios,this paper proposes a semantic segmentation method of nighttime obstacle image based on attention mechanism.The main research contents are as follows:(1)Aiming at the problem that the existing obstacle segmentation data sets are based on the daytime visible scene,while the nighttime scene data sets are relatively few,this paper constructed a near infrared nighttime road obstacle image data set.Firstly,the near infrared camera with light filling device is selected as the acquisition equipment of image data set.Secondly,image preprocessing is performed on the collected images.Then,annotate the image and expand the data set;Finally,statistical analysis,classification imbalance correction and data set division are carried out to construct the near infrared nighttime road obstacle image data set.(2)Aiming at the problem that the platform of obstacle segmentation method will directly affect the segmentation accuracy and speed of obstacles on the driving road,the conventional and classical image semantic segmentation network models FCN,Seg Net,UNet,PSPNet and DeepLabV3+ are conducted experimental comprehensive comparative analysis on Cam Vid universal data set.The results show that DeepLabV3+ has the best comprehensive segmentation performance.Therefore,DeepLabV3+ network model is selected as the nighttime obstacle image segmentation platform in this paper.(3)Aiming at the problem of rough edge segmentation in DeepLabV3+,an Attention-DeepLabV3+ nighttime road obstacle image segmentation method is proposed and experimental research is carried out in this paper.Firstly,the principle of attention mechanism is elaborated and the advantages and disadvantages of the two kinds of attention mechanism and the applicable scenarios are analyzed theoretically.Then,the channel attention mechanism and DeepLabV3+ network model are fused,and the Attention-Deep Labv3+ network model algorithm is proposed.Finally,the algorithm is applied to City Scapes,PASCAL VOC2012,as well as the self-built near infrared nighttime road scene dataset.The results show that the segmentation accuracy of the proposed algorithm in these three datasets is 85.7%,90.2% and 82.1%,respectively.3.6,1.2 and 3.5 percentage points higher than the original DeepLabV3+;There is little change in the segmentation speed,which can be almost ignored.To a large extent,the superiority,effectiveness,robustness and applicability of the improved algorithm are verified.