Road Traffic Information Detection And Recognition Based On Vehicle Video

In recent years,autonomous driving and driving assistance systems have become a research hotspot.In automatic driving and vehicle navigation systems,road traffic information data is one of the most important auxiliary data,which makes the detection and recognition technology of road traffic information based on in-vehicle video have important significance and research value.The traditional road traffic information detection and recognition methods are mainly based on edge,texture feature detection or image segmentation.Due to the complexity and change of the driving environment,the traditional method has poor robustness.In recent years,deep learning has made tremendous achievements in the field of computer vision.In this paper,based on deep learning technology,the road line detection and road marking detection and recognition in road traffic information are studied.The specific research content is as follows.(1)The road lane detection based on deep learning image segmentation algorithm is studied.The VGG16 model with self-attention distillation structure is used as the basic network of the segmentation network;the Bi-DU structure is proposed as the head of the segmentation network;finally,the Cross Max Pool thinning algorithm,DBSCAN clustering algorithm and polynomial fitting algorithm are used to get the center line of the lane.A precision rate of 0.9661 and a false detection rate of 0.0294 and a missed detection rate of 0.0249 were obtained on the tusimple data set.The experimental results show that the lane lines extracted by this method are more complete,and compared with the traditional method,a significant accuracy improvement is achieved.(2)The detection and recognition of road marking based on deep learning object detection algorithm are studied.Anchor-free mode object detection architecture is proposed to detect and recognize road signs.To facilitate the integration of subsequent models,the basic network adopts the same as the lane segmentation algorithm,and the network finally outputs the probability map of object center point,bounding box and object center point offset.The algorithm achieves a detection accuracy of 0.828 mAP on the road marking data set collected in this paper.The experiment proves that the detection and recognition algorithm of road markings proposed in this paper has achieved high accuracy,compared with the traditional object detection method,it is more robust and has certain practical application value.(3)The multi-task learning and model compression of deep learning are studied.The lane line detection model and the road marking detection and recognition model share a basic network to achieve model fusion,and the two detection modules are alternately trained during the training process;then the model pruning technology based on the γ parameter in the BN layer is used for compression.The accuracy of the original model is achieved on the lane line and road marking data set,but the detection speed is nearly 4 times faster and the size of model decreases 82.5%.Experimental results show that the introduction of multi-task learning and model compression can speed up the model detection speed while ensuring accuracy,thereby meeting the performance requirements for real-time detection in practical applications.Experimental results show that the method proposed in this paper can detect lane lines and markings on the road from vehicle video.Thanks to multi-task learning and model compression technology,the lane line and road marking detection and recognition models can achieve faster speed and higher accuracy at the same time,which has certain practical application value.