Occluded Area Recognition And Cooperative Perception For Intelligent Vehicles

In the past decade,with the development of the transportation industry and the booming national economy,China’s automobile ownership has grown rapidly,and the total traffic mileage has experienced explosive growth.The significant increase in the number of cars has brought great convenience to people’s lives and travel,but it has also inevitably led to an increase in the number of traffic accidents,causing enormous loss of life and property to the country and its people.In order to improve the driving safety of motor vehicles,reduce the possibility of accidents caused by improper driving operations by human drivers,and improve road traffic efficiency,intelligent vehicle technology has emerged.Compared to traditionally driven vehicles,intelligent vehicles use onboard sensors to perceive the surrounding driving environment before making rational driving behavior decisions.They can quickly plan future driving trajectories and accurately execute corresponding control operations autonomously.Therefore,enabling intelligent vehicles to achieve accurate and comprehensive environmental perception is the first step towards ensuring driving safety and moving towards fully autonomous driving.Comprehensive environmental perception is the key to intelligent vehicles.However,current intelligent vehicle perception algorithms only recognize targets within the intelligent vehicle’s field of view as much as possible,without identifying heavily or completely occluded background targets or predicting/annotating occluded areas caused by foreground objects.This approach limits the comprehensive perception and understanding of the current driving environment by intelligent driving systems.In this paper,a semantic segmentation model is designed,which takes the image data captured by the intelligent vehicle’s surround cameras as input.By using multi-source information fusion technology and spatial transformation networks,it accomplishes viewpoint transformation tasks and extracts image features using a dilated spatial convolutional pooling pyramid structure based on inverted residual structures.By using dilated convolutions for upsampling the feature maps,a lightweight EncoderDecoder semantic segmentation model suitable for intelligent vehicle scenes is constructed.This model can output end-to-end semantic segmentation perception results of the intelligent vehicle’s driving environment,including occluded areas,from a bird’s-eye view.Furthermore,this paper does not rely on manually labeled data but collects the dataset through the Carla simulator and automatically completes the subsequent data annotation using the designed ray-localization method.Through experimental verification on the collected dataset,the proposed method achieves a Mean Intersection over Union(MIoU)score of 71.49%,surpassing traditional models based on perspective transformation principles and fully connected networks for viewpoint transformation.Cooperative perception is a technology that utilizes communication technology to enable intelligent vehicles to break through the perception limits and information isolation of single-vehicle perception technology,improve the reuse rate of perception information between vehicles or between vehicles and the road,and reduce the manufacturing cost of single vehicles.In this chapter,a vehicle-vehicle cooperative perception method is proposed,which mainly addresses two problems in intelligent vehicle cooperative perception technology:"whether cooperative perception is needed" and "how to achieve cooperative perception." The proposed method takes the perception results of a single intelligent vehicle from a bird’s-eye view as input,accurately locates the areas requiring cooperation by determining the location of occluded areas and using a quadtree region division method.It calculates the contribution of surrounding intelligent vehicles to the areas requiring cooperation based on the designed correlation calculation method,selects suitable intelligent vehicles as providers of cooperative information,and initiates subsequent cooperative information requests.Finally,it integrates all obtained cooperative information at the result level.Thanks to the design of the "query,request,communication" three-stage method,the proposed method reduces the number of communication link establishments and the amount of cooperative information transmission compared to existing methods while ensuring the effectiveness of cooperative perception.Through comparative experimental results on the V2X-Sim dataset,the proposed method achieves higher Communication Improvement Scores(CIS)for vehicle and pedestrian semantic objects compared to other existing cooperative perception methods.