Design And Implementation Of A Multimodal Data Selection System For Autonomous Driving

With the development of autonomous driving technology,the types of sensors deployed on self-driving vehicles are gradually increasing.Due to the different characteristics of these sensors,the quality of their data in different environments is different.In order to provide high quality sensory data to the upper layer applications,it is necessary to evaluate and optimize the quality of the data collected by different sensors.Some researchers have proposed using target recognition accuracy as a method for data quality assessment.This method selects different modal data for applications such as target recognition,and compares the quality of different modal data based on the application results to select the better result.However,this method needs to calculate and process multiple dimensional data first,which is relatively inefficient to use.To this end,this paper proposes a multimodal data quality assessment model and designs and implements a multimodal data selection system for autonomous driving in combination with this model.The system first analyzes data quality in dynamic scenes based on the quality assessment model and selects high-quality modal data based on the data quality score before subsequent data processing and computation,avoiding unnecessary computational overhead.The main work of this paper is as follows:(1)This paper proposes a multimodal data quality evaluation model,which sets a variety of data quality evaluation indexes that are compatible with multiple modalities including camera and LiDAR,and performs data quality evaluation and comprehensive scoring for multiple modal data collected by autonomous driving sensors in dynamic scenes,so as to provide quality selection basis for upper layer application services.(2)Based on the proposed multimodal data quality evaluation model,this paper designs and implements a multimodal data quality selection system for autonomous driving.According to the system development process of software engineering,requirements analysis,architecture design,module division,system implementation according to the design,and functional and non-functional tests were conducted.The tests show that the system designed and implemented in this paper has data selection capability for multimodal data.