Road Scene Detection Method Based On Deep Learning

Road scene detection is a key technology for the implementation and application of assisted driving and autonomous driving.Not only that,road scene detection can also detect the driver’s blind spots and reduce the incidence of traffic accidents.Two algorithms,semantic segmentation and target detection,can be used to detect on demand.For lane lines,driveways,crosswalks,buildings and other targets with large areas and wide ranges,semantic segmentation segments different classes of targets from pixels,which is difficult to be compared with other methods.Target detection,on the other hand,allows for better detection of pedestrians,vehicles,traffic lights,traffic signs,etc.in road scenes from the instance level.A bilateral U-Net network model incorporating a spatial attention mechanism is designed to address the problems of high cost and time of dataset production and the poor effectiveness of existing semantic segmentation models for segmenting small sample imbalanced datasets.The model uses lightweight Mobile Net V2 as the backbone network for feature hierarchical extraction,and proposes a null spatial pyramidal attention(APSA)module in contrast to the null spatial pyramidal(ASPP)module,which can increase the perceptual field and enhance information,and finally adds a contextual fusion prediction branch to fuse high and low semantic prediction results to help detail and edge pixel correctly classified.The model achieves an m IOU of 75.85% on the Cam Vid dataset.And,to verify the generalizability of the model and the usefulness of the APSA module,experiments were conducted on the VOC 2012 dataset,and the APSA module boosted the m IOU by about 12.2%.The experiments demonstrate that the model can achieve the desired road segmentation effect with a small number of datasets.A lightweight Center Net model with fused context based on Center Net is proposed for the problems of many parameters of target detection algorithm,large computation and poor fusion of contextual information.The lightweight Mobile Net V3 network is used as the backbone network,and the SE module in it is replaced by the ECA module to achieve the reduction of the number of parameters and the model.Then,according to the characteristics of the target detection algorithm,the APSA module is improved to expand the convolutional perceptual field by increasing the number of voids,and it is fused in the level-by-level connected contextual feature fusion module to better fuse the contextual information while reducing the number of consecutive upsampling,thus improving the detection accuracy of the network.Compared with Center Net,the detection speed FPS is improved by about 40% and the model is reduced by about two-thirds with only 3.2% lower accuracy,so the detection effect in practice does not lose the original model.It has more practical application value.Through experiments,it is proved that the semantic segmentation method proposed in this paper can improve the performance of road scene segmentation for small data sets,and solve the problem of time-consuming and laborious data set production to a certain extent while the accuracy is steadily improved.The proposed target detection algorithm takes into account lightweighting and information fusion to improve the detection speed while ensuring the detection accuracy.Both methods can be applied in practical engineering and provide ideas for subsequent research.