Research On Detection System Of Road Water Accumulation For Vehicle Active Safety

When the vehicle is driving on uneven water accumulation road surface,it is easy to cause dangerous phenomena such as sideslip and deviation.when it reaches the critical vehicle speed,due to the hydroplaning,the wheels are easy to get out of control,thus causing traffic accidents.Therefore,the detection of water covering the road is very important for the safety of highspeed automatic driving vehicles in complex environment.At present,the detection technology of water-road condition is mainly used to qualitatively judge the road condition,which can’t further obtain the distribution information of water covering the road.And it is difficult to meet the demand of automatic driving under complex road conditions.Therefore,based on the National Key R&D Program of China: Development and Application of New Multi-functional Intelligent Vehicle Terminal(2018YFB1600701),this paper focuses on the road surface water state recognition method based on transfer learning and the method of segmentation of pavement ponding area based on depth supervision,and designs the road surface water state detection system based on embedded GPU platform.Firstly,the road surface water state recognition method based on manual extraction of image features is easy to be interfered by the external environment,resulting in poor generalization performance of recognition,and the recognition method based on deep learning has high requirements for data size.A road surface water state recognition method based on transfer learning is proposed.Taking Image Net dataset as the source data,the convolution layers of Inception V3 are used for feature migration under isomorphic data,and then the new road state classification layer is connected to establish the road surface water state recognition model.Through the field acquisition and network search,a large number of different degrees of water accumulation images are acquired to establish the pavement database.And the model training and test are completed.The experimental results show that the model constructed in this paper has higher recognition accuracy than other similar recognition models.Secondly,aiming at the problem that the road surface water has no fixed shape and has mirror reflection interference,the Full-Resolution Residual Networks(FRRN)is selected as the segmentation network framework through the comparative analysis of common semantic segmentation networks.Because the loss function of FRRN can’t take into account the global pixels,and can’t provide direct supervision for the training of low and middle level network.By adding 1 × 1 convolution and dice loss to the four decoding layers of FRRN,a segmentation model of pavement ponding area based on DSFRRN(deep supervised FRRN)is established,which can reduce the loss of effective feature information and improve the nonlinear expression ability of the model.Through labeling a large number of ponding images,the ponding segmentation dataset is established.And the 5-fold cross validation method is used to train and test the model.Compared with the original model and similar segmentation model,the improved model has the best segmentation performance.Then,the hardware architecture of road surface water state detection system is designed,and the selection of hardware equipment such as image acquisition and image processing platform is completed.According to the characteristics of hardware platform,GPU parallel processing and neural network model calculation are combined to speed up the prediction speed of the system.Configure Jetson TX2 software environment to provide support for subsequent program migration.Finally,in order to verify the robustness and real-time of the system,the system program and model weight are transplanted to the Jetson TX2 embedded platform,and the detection equipment is built to detect the water state of asphalt pavement under various weather conditions.The experimental results show that the system has high accuracy and stability of road water state recognition,good segmentation performance and robustness of ponding area,and real-time performance can meet the demand of roadside unit.