Cross-weather Road Scene Re-Identification

With continuous development of science and technology,autonomous driving technology is constantly improving.In the near future,autonomous driving vehicles will appear in daily life.For autonomous driving,road detection technology is the key,which includes detection of road extension direction,road marking and the detection of obstacles in the road.At present,road detection is mainly carried out by vehicle camera,infrared detector and so on.However,in daily life,vehicles often pass through the roads with insufficient illumination and bad weather condition.In these cases,traditional road detection technology will encounter great challenges.Therefore,it is very important to design and implement a road detection system with strong robustness to illumination transformation and different weather conditions.Considering daily driving is highly reproducible and change of road conditions is small over a short period,this paper proposed a method that matches a road image with poor lighting conditions and poor weather conditions with a same point road image which is taken in the same position but with sufficient lighting and good weather conditions.Then detect the road with the latter image.In this way,the system is robust to different environmental conditions such as change of illumination and weather.The main research contents are as follows:(1)Due to the uniqueness of the subject,set up the road scene image dataset.The dataset was manually marked with two kinds of illumination conditions(daytime and nighttime)and five kinds of weather conditions(sunny,snowy,rainy,foggy and cloudy day).The number of images in the dataset is close to ten thousand.Besides,in the dataset,there are 12,387 pairs of manually marked images taken at the same point but with different environmental conditions.The dataset can be used in the research of road image matching algorithm and road related algorithms.(2)This paper uses AlexNet deep neural model as title basic model.In order to solve this problem,this paper took experiments to study the performance of the feature matrices output by different layers and the performance of the network pre-trained by different datasets.Finally,this paper determined the feature of fifth layer,used the AlexNet network model which was pre-trained by ImageNet and Places as the basic model,and used the CMC curve as the standard function of the performance comparison.(3)Designed and implemented the cross-weather road scene re-identification system.All the source domains below were night and rainy night.The parameters of the basic network model were fine-tuned to improve the system’s performance,and the average Rank 1 increased by 1.7%and 8%respectively.Further,after performing the metric learning,the deep feature was transformed and dimension was reduced to a smaller common subspace,then the images were retrieved and matched.After comparison,the average Rank 1 incresed by 3%and 10.67%resprectively.What’s more,the performance based on the SCNN network training model was compared with that of the feature transform.As result,the average Rankl was increased by 1.5%and 2.5%respectively.Finally,it was shown that the forward propagation speed was 517.3 fps/s,the overall training speed was 263.2 fps/s,and the image processing speed was 17.3 fps/s in practice,which met the practical requirements.