基于卷积神经网络感知功能的极端场景生成方法

Corner Cases Generation for Virtual Scenario-based Testing of CNN-based Autonomous Driving Function

摘要:

近年来，基于卷积神经网络深度学习的感知算法在自动驾驶车辆环境感知系统中发挥着越来越重要的作用。由于在神经网络训练过程中，训练数据无法覆盖所有极端场景，因此如何保证基于深度学习的感知算法在极端场景下的安全性和可靠性，仍是一个亟待解决的问题。传统的基于真实行驶里程的验证方法，在获取极端场景数据上危险性高，经济性差，因此很难检验驾驶功能在极端场景下的性能。基于虚拟场景的仿真验证方法，虽然可以通过设置场景参数来生成大量测试场景，但是通过简单的参数组合并不能有效的生成极端场景。本文展示了一种在虚拟环境中生成极端场景的方法，用于训练和测试基于深度卷积神经网络的车道线识别算法。首先将场景特征用参数进行表示，然后使用deep Q-learning强化学习的方法，来生成极端场景的参数组合。通过与随机组合以及成对组合场景参数的方法进行对比，可以看出该基于强化学习的场景生成方法可以更有效地生成极端场景，因此可提高自动驾驶感知功能的测试效率，同时可为卷积神经网络提供更多的极端场景训练数据。

Abstract:

Deep learning-based perception algorithms have gained importance in autonomous vehicle perception systems in recent years. Since the training data cannot cover all critical scenarios and corner cases， how to ensure the safety and reliability of deep learning-based perception functions in crucial scenarios is still an open challenge. Conventional approaches test the driving functions in real-life environments， which can be risky and uneconomic to validate in corner cases. Virtual scenario-based simulation validation approaches can generate a large number of test cases by setting test scenario parameters， but the purely combinatorial parameter cannot effectively generate corner cases. In this paper， we present a novel approach to generating corner cases in a virtual environment for validation of a CNN （Convolutional Neural Network）-based lane detection function. We represent the scene features with parameters， and then use the deep Q-learning reinforcement learning approach to generate the parameter combinations of corner cases. In addition， by comparing with the approaches of random combination and pairwise combination of scene parameters， our approach can generate corner cases more efficiently and improve the testing efficiency of the autonomous driving perception functions.