辅助驾驶系统中浓雾天识别方法分析

辅助驾驶系统中通过计算能见度信息对雾天进行识别的方法存在一定的局限性。针对此问题,利用机器学习方法设计了可用于辅助驾驶系统的浓雾天识别算法,避免了基于能见度识别雾天的局限性。首先建立了基于驾驶场景的图像训练集;然后基于卷积神经网络和胶囊网络分别设计卷积神经网络-浓雾识别(convolutional neural networks-dense fog recognition, CNN-DFR10)和胶囊网络-浓雾识别(capsule networks-dense fog recognition, CN-DFR5)两种浓雾天识别算法模型,算法模型通过设置概率阈值的方法对天气类型进行区分,不同的概率值范围对应不同的天气类型。最后对比分析CNN-DFR10和CN-DFR5在浓雾天、雨天、阴天和晴朗天4种天气类型中的测试结果。结果表明:CNN-DFR10算法对天气的识别准确率为86.9%,CN-DFR5算法的识别准确率为97.5%,后一种算法比前者能够更有效地从4种天气类型图像中区分出浓雾天和非浓雾天。

Analysis of Recognition Method of Dense Fog in Assistant Driving System

In assisted driving systems, the method of using visibility information to identify foggy days has certain limitations. Aiming at this problem, the machine learning method is used to design a dense fog recognition algorithm that can be used in assisted driving systems, avoided the limitations of the visibility method. Firstly, an image training set based on driving scene was established. Then CNN-DFR10 and CN-DFR5 algorithm models was designed based on the convolution neural network and capsule network, respectively. The algorithm models distinguished weather types by setting probability thresholds, and different probability value ranges corresponded to different weather types. Finally, by comparing and analyzing the test results of CNN-DFR10 and CN-DFR5 in the four weather types of dense fog, rain, cloudy and sunny, it is found that the accuracy of the CNN-DFR10 algorithm for weather recognition is 86.9%, the accuracy of the CN-DFR5 algorithm is 97.5%. The latter algorithm is more effective than the former in distinguishing dense fog and non-dense fog from the four types of weather images.