Research On Real-time Semantic Segmentation Of Autonomous Driving Based On Improved Dense ASPP

With the rapid development of computer vision and automotive intelligent technology,the era of autonomous driving has arrived.Recognition,perception and understanding of the surrounding environment are indispensable technology to develop autonomous driving.As an important means of scene understanding,semantic segmentation technology has become a solution to this problem.The images in the autonomous driving scene have the characteristics of high resolution,large scale changes and high running speed requirements,which poses a great challenge to the accuracy and real-time performance of current image semantic segmentation technology.In response to the requirements of high real-time and high accuracy in autonomous driving scenarios,this paper studies and improves the Dense Atrous Spatial Pyramid Pooling(DenseASPP)method in a multi-scale feature extraction model,and proposes an improved DenseASPP(IDenseASPP)Real-time semantic segmentation method.The main studying contents of this paper are as follows:(1)To improve the processing speed,this paper proposes a lightweight and fast downsampling strategy.It introduces deep separable convolutions to lighten and improve the structure of shallow convolutional networks,and achieves fast extraction of small-scale feature maps.Compared with mainstream lightweight backbone networks,generating feature maps of the same size,the strategy has improved more than doubled in speed.At the same time,based on this strategy,this paper applies the complex modules in IDenseASPP to tiny feature maps,which effectively balancing the contradiction between speed and accuracy,and improving the whole operating speed.(2)To improve the accuracy of segmentation and alleviate the problem of accuracy degradation caused by tiny feature maps,this paper introduces a hybrid dilated convolution algorithm into the DenseASPP module and constructs the IDenseASPP module.This module can build denser multi-scale feature pyramids with fewer parameters,and alleviate the grid problem unique to dilated convolutional combinations at the same time.Experimental results show that compared with the traditional multi-scale feature extraction module and densely connected module,the IDenseASPP module brings an accuracy improvement of more than 1.6% in the mean Intersection over Union(mIoU)metric.(3)This paper studies the effect of the input image size on the receptive field center of the output pixel of the upsampling layer in the coder-decoder structure.Control experiments show that the use of an optimization scheme combining align corners and a specific input image size can effectively overcome the skewing problem of the receptive field center,and thus bring about an improvement in accuracy.Combining the above strategies with other skills,such as low-level feature selective cascading and multi-size label supervised training,this paper trains and evaluates the IDenseASPP method on two datasets,Cityscapes and CamVid.Finally,the segmentation accuracy of IDenseASPP on the test sets of the two data sets reached 72.4% mIoU and 69.2% mIoU,and the real-time segmentation frame rates on the NVIDIA GTX 1080 Ti GPU reached 113 FPS and 156 FPS.The results verified the feasibility and effectiveness of the method.In addition,this paper applies IDenseASPP to the problem of drivable area detection.Training and experiments of drivable area detection were performed on the Kitti-Road and Cityscapes datasets,and real scene tests were performed on the intelligent driving experimental platform,which further verified the practicability of the method.