微型电脑应用

3D目标检测是计算机视觉的研究重点。目前，3D目标检测模型的精度不断提升，但模型的鲁棒性较低，存在缺乏测试预言的问题。为了解决测试预言问题，提出面向3D目标检测模型的蜕变测试技术。模拟现实场景中的噪声干扰和雷达校准问题；提出增加噪声、删除部分点、角度畸变和径向缩放4种蜕变关系；构建后续点云生成方法，设计点云的一致性验证方法，以判断原始输出和后续输出是否满足蜕变关系。对3个先进的3D目标检测模型进行实验验证，结果表明，所提出的方法能够检测出大量3D目标检测模型的错误行为，4种蜕变关系识别错误行为的平均错误率分别为18.46%、9.93%、17.10%和18.99%，能够有效地评估模型的鲁棒性。

3D object detection is a key research focus in computer vision.Currently,the accuracy of 3D object detection models is continuously improving,but robustness of the models remains low.There is a lack of testing oracles.To solve the issue of testing oracles,metamorphic testing techniques for 3D object detection model is proposed.Noise interference and radar calibration issues in real-world scenarios are simulated,and four metamorphic relations are proposed,i.e.,adding noise,deleting partial points,angle distortion and radial scaling.A subsequent point cloud generation method is constructed,and a consistency validation method for point clouds is designed to determine whether the original output and subsequent output satisfy the metamorphic relations.Experimental verification is conducted on three advanced 3D object detection models.The results show that the proposed method can detect a large number of erroneous behaviors in 3D object detection models.The average error rates of the four metamorphic relations identifying erroneous behaviors are 18.46%,9.93%,17.10% and 18.99%respectively,which can effectively evaluate the robustness of the models.