Research On Single-stage Instance Segmentation Algorithm Based On Convolutional Neural Network

With the rise of a new round of scientific and technological revolution and industrial transformation,intelligent driving vehicles have become the strategic direction of the Global automotive industry development.Intelligent driving vehicles is a comprehensive system integrating environment perception,intelligent decision-making,and control execution,which has received extensive attention and research in recent years.Real-time and accurate environmental perception is the premise and foundation for the safe and reliable driving of intelligent driving vehicles.Image instance segmentation technology is widely used in the field of intelligent driving perception.Currently,it is mainly used to identify lane lines and other necessary traffic information,or to match the segmentation results with the point cloud data of laser sensors for achieving pixel-level multi-sensors fusion.Deep learning model training requires large-scale data sets as support.However,due to the time-consuming and labor-intensive labeling of instance segmentation data sets,there are still few data sets used for instance segmentation of intelligent driving traffic scenes.At the same time,existing single-stage segmentation algorithm can meet a certain real-time performance,it brings is the decline of accuracy.But the intelligent driving traffic scene is complex and changeable,the instance segmentation of small objects generally lacks details,so the segmentation accuracy of the algorithm is still very low.In addition,the computing power of intelligent driving in-vehicle equipment is low,but the current intelligent driving instance segmentation algorithm relies on powerful computing resources,which is one of the obstacles to the large-scale popularization of intelligent driving systems.In view of the above problems,this paper constructs an instance segmentation dataset of intelligent driving traffic scenes.Based on the single-stage instance segmentation Yolact algorithm,the path aggregation feature pyramid module and convolution block attention mechanism are introduced to optimize the algorithm to improve the segmentation accuracy of the algorithm.;Finally,the network lightweight design is carried out on the algorithm to improve the real-time performance of the algorithm.The details are as follows:(1)Construction of intelligent driving traffic scene instance segmentation dataset and data enhancement.Because the labeling of instance segmentation datasets requires a lot of time and effort,currently instance segmentation datasets for intelligent driving traffic scenarios are very scarce.In this paper,manual instance-level annotation is performed based on the KITTI dataset images,and finally a basic instance segmentation dataset is generated.In order to improve the generalization of the network and reduce over-fitting,the data enhancement methods of single data deformation,random image stitching,and random instance mask clipping are used to enhance the data of the basic instance segmentation dataset.Experiments show that data augmentation can effectively improve the training effect of the algorithm model,and finally the instance segmentation data set used in this paper is obtained.(2)Instance segmentation algorithm optimization based on path aggregation feature pyramid and convolution block attention mechanism.In view of the low segmentation accuracy of the single-stage instance segmentation algorithm and the difficulty of small target segmentation,this paper is based on the feature pyramid of the Yolact algorithm,considering that the shallow network information contains a large number of edge shapes and other features,which are very useful for instance segmentation tasks.Path aggregation feature pyramid is introduced to shorten the propagation path of shallow feature information.In addition,the convolution block attention mechanism is used in the backbone network of the algorithm to improve the effect of the small object segmentation.(3)Lightweight of the instance segmentation algorithm.Most of the intelligent driving in-vehicle devices are mobile devices with low computing power.In view of the problem that the Yolact instance segmentation algorithm still has a large amount of network parameters and computation,and the inference speed is slow,this paper uses the current mainstream lightweight network to improve the Yolact algorithm;Then compare the improved algorithm with the Yolact benchmark algorithm,this experiment shows that the lightweight scheme reduces the amount of parameters and calculation of algorithm,and improves the real-time performance of the Yolact algorithm.