Research On Instance Segmentation Algorithm Of City Street Scenes Based On Deep Learning

With the popularization of automobiles and the improvement of people’s living standards,as well as the rapid development of artificial intelligence and computer vision technology,smart cars and autonomous driving have become new demands and new expectations for people’s future life.As one of the components of an autonomous driving system,the perception module provides important information for subsequent decision-making and control modules,and its environmental perception effect on the ubiquitous urban street scene becomes a very critical issue.Instance segmentation as a technology that is more in line with human perception of the environment can help the perception module better establish an understanding of the scene.In this paper,based on deep learning,an instance segmentation algorithm for city street scenes is studiedFirst of all,this paper designs a lightweight network structure for a two-stage instance segmentation network.Structures such as depthwise separable convolution,inverted residuals block with linear bottleneck and SE block are introduced through the three networks of MobileNetV1,MobileNetV2 and MobileNetV3 respectively.Each backbone is combined with the feature pyramid FPN to fuse multi-scale information as the feature extraction part of the two-stage instance segmentation algorithm.Secondly,in view of the lack of individual categories in the dataset,this paper improves the feature extraction ability of the network through pre-training of the COCO dataset.Finally,based on different algorithmic schemes,this paper improves non-maximum suppression,using two methods,Soft-NMS and Fast-NMS,respectively.Soft-NMS introduces a mechanism to reduce the confidence of BBox on the basis of traditional non-maximum suppression,while Fast-NMS improves the traditional non-maximum suppression through matrix calculation and relaxation of screening mechanismIn this paper,we use city streetscape images in the Cityscapes dataset to train and verify the algorithm,and evaluate the accuracy,speed,and visualization of each algorithm Through experiments,the combination of MobileNetV3 and FPN,with the pre-training on the COCO dataset obtained slightly higher accuracy than Mask R-CNN,and on the basis of it,the inference speed of the model was improved,and the parameters and computational complexity of the model were reduced.And on the basis of not increasing the amount of network parameters and training difficulty,Soft-NMS and Fast-NMS respectively improve the accuracy of the model and the speed of inference.