面向自动驾驶感知的快速不确定性估计方法

A Fast Uncertainty Estimation Method for Autonomous Driving Perception

在自动驾驶的视觉感知任务中,准确且快速提取认知不确定性和偶然不确定性对有效解决自动驾驶的预期功能安全问题至关重要。传统方法中,如Monte-Carlo Dropout和Deep Ensembles,通过采样不同子模型的预测结果来估计不确定性,这使在模型推理阶段不确定性估计速度很慢且容易占用处理器大量内存。针对Monte-Carlo Dropout不确定性估计速度较慢及其后续检测结果选取的问题,提出了一种快速Monte-Carlo Dropout方法及后续检测结果校正的方法。此方法使用多头机制替换了Monte-Carlo Dropout传统的多次采样机制,节省了采样时间,进而节省了整个不确定性估计阶段的推理时间。

In the visual perception task of autonomous driving,it is crucial to accurately and quickly extract the cognitive and accidental uncertainties to effectively resolve the Safety of the Intended Functionality(SOTIF)issues associated with autonomous driving.In traditional methods such as Monte Carlo dropout and deep ensembles,uncertainty is estimated by sampling the prediction results of different sub-models,which slows down the estimation and tends to occupy a large amount of memory in the processor during the model inference stage.A fast Monte Carlo dropout method and a technique for correcting subsequent detection results are proposed to address the issues of slow estimation of uncertainty in Monte Carlo dropout and the selection of subsequent detection results.This method uses a multi-head mechanism to replace the traditional multiple sampling mechanism in Monte Carlo dropout,thereby saving time in both sampling and inference throughout the uncertainty estimation process.