Design And Implementation Of Highway Condition Recognition And Detection System

With the development of artificial intelligence,intelligent driving has become a hot research field at present,and road condition detection is a key part of this field.At present,most intelligent driving cars use a combination of sensors to detect road conditions.However,this method has high cost,large amount of data and complex processing.Therefore,in recent years,pure visual deep learning methods have been gradually used in the research of road condition detection.After analyzing the research direction and key technologies of road condition detection,this thesis decides to adopt pure vision deep learning method for road condition detection,deeply studies the road target detection algorithm,road semantic segmentation and multi-task road detection algorithm,proposes several improved algorithms,and finally realizes the road condition detection system according to the algorithms.The main work of this thesis is as follows:(1)The current road condition detection algorithm does not pay much attention to the road condition of rural roads.In order to solve this problem,two road target detection networks are trained in this thesis.The first is the urban highway model,aiming at various vehicles and traffic signs on the urban highway;The second is a rural road model that focuses specifically on cars,pedestrians,motorcycles,and livestock targets on rural roads.After training and verification,both of the two networks have achieved a high detection effect.(2)An improved highway target detection algorithm based on Yolov5 algorithm was proposed through in-depth study of the target detection algorithm.smooth-FReLU activation function was introduced to solve the problem of low spatial sensitivity of activation function in Yolov5.The Ghost convolutional structure and Rep GFPN structure are used to improve the Yolov5 network,which is complicated and has insufficient feature integration.α-IoU and focal-EIoU loss functions were introduced to optimize the loss function and reduce the influence of sample imbalance.A variety of image enhancement training methods are adopted and integrated.Finally,the experimental results show that the detection effect of the improved algorithm is significantly improved compared with the Yolov5 algorithm.(3)Aiming at the problem that multiple road condition detection tasks consume a lot of resources at the same time and the algorithm is too complex,a lightweight multi-task detection algorithm is proposed to solve three road condition detection tasks:road target detection,lane line segmentation and driveable area segmentation.Using the improved backbone network as the encoder,two branch decoders of split task are constructed by step-up sampling method.The final experiment shows that on the BDD100K data set,the multi-task algorithm proposed in this thesis is relatively superior,and the detection speed is improved significantly.(4)With multi-task algorithm as the core to realize the highway condition detection system,two road condition modes of urban highway and rural highway can be selected for picture,video and real-time detection,providing functions such as driving area segmentation,lane marking and target detection.It has practical application value to detect and remind the relative safe distance of the target in real time,and finally assist the driver to complete the correct decision in the driving process.