Automatic Driving Image Data Augmentation Technology Based On Functional Safety

With the development of artificial intelligence technology and sensor technology,autonomous driving technology has been widely used.But the safety implications of autonomous driving have also become a concern.Accidents caused by self-driving cars have happened frequently in recent years.Investigations later revealed that most of these accidents were caused by defects in the system that were not detected during test-ing.Therefore,different data augmentation methods are proposed to test it.Although the existing methods can quickly generate a large number of image data of different driving scenes,the data types obtained in this way are relatively single.To solve this problem,we propose an automatic driving image data augmentation technology based on functional safety.This paper combines the functional safety analysis technology in the automotive industry development and testing field with the image augmentation technology in the deep learning testing field,and uses the image processing method to simulate the image data when various abnormal situations occur in the image data acquisition path of the autonomous driving vehicle.In this paper,the failure mode and impact analysis method(FMEA)recommended in ISO26262 functional safety standard is adopted to slice and analyze the overall in-put process of the image data acquisition system of autonomous driving vehicles,and the possible failures and abnormal failure forms in each part of the image acquisition process are classified and summarized.Summarizes the abnormal types that are most likely to fail in the image acquisition system.Secondly,this paper analyzes the abnor-mal types that are most prone to failure in the image data flow path,analyzes the reason and mechanism of producing abnormal image data,as well as the impact and changes of these failures on the image data.According to these reasons and mechanisms,this paper summarizes the rules of data augmentation when these anomalies occur,and uses Python language and Open CV library to implement the augmentation data.Finally,the modified KITTI data set was used to train the target recognition model,and the test data of the KITTI data set were randomly selected for data augmentation.We carried out target recognition test on these augmented image data,observed the test results of the target recognition model,and analyzed and summarized the impact of the augmented image on the deep learning model.After testing the augmented data,it was found that the image data generated by the functional safety-based autonomous driving image data augmentation technology would have varying degrees of impact on the judgment of the deep learning system.These image data can lead to misdetection and missing detection of target detection and the decrease of the confidence of target detection.The images generated by this technology can detect the deficiencies and loopholes in the deep learning system when abnormal situations of autonomous vehicles occur,thus reflecting the security and ro-bustness problems within the system.