Research On Road Turning Sign Detection Algorithm Based On Deep Learning

In recent years,with the rapid development of intelligent vehicle industry,automatic driving and intelligent assisted driving have become the research hotspot of researchers.The core of autonomous driving technology is to capture real-time road conditions and identify the surrounding road environment through equipment such as on-board cameras,laser rangefinders and radars.Ground traffic sign detection is one of the key technologies in autonomous driving research,mainly including lane line recognition and turning sign recognition.Among them,lane line recognition technology has become mature,and the recognition of turning signs is mostly based on traditional image processing methods.Such methods are less robust and cannot cope with special situations such as sudden changes in light,shade occlusion,and weather changes and so on,and the detection speed cannot meet the needs of real-time detection.In addition,due to the strong perspective distortion in the road live map taken by the on-board camera,the long-distance target presents a small and fuzzy state,resulting in missed detection and wrong detection in the detection process.In order to realize the accurate recognition of road turning signs,this thesis proposes a road turning sign recognition algorithm based on deep learning and perspective down-sampling.The algorithm can adaptively learn the target features without excessive human intervention and can improve the detection accuracy.The main contents of this thesis are as follows:1.In view of the current lack of road turning sign data sets,a multi-scene data set containing 22,000 images is created to complete the data labeling work;using this data set,various classic algorithms such as YOLO series algorithms,Faster R-CNN and SSD are compared.For comparison experiments,YOLOv3-tiny is selected as the basic network of sign detection,and the anchor size,output scale of prediction module and network structure of the algorithm are optimized.The improved network is tested on self-made data sets.The average image time is reduced from 2.15 milliseconds in the basic network to 1.95 milliseconds,and the mean average precision value is increased from 79.37%to 92.05%.2.In the test results,there are some problems such as missing detection of small targets and low detection accuracy caused by image perspective distortion,this thesis proposes a road image processing algorithm based on perspective down-sampling,selects the area of interest containing small targets,and removes complex irrelevant background information;Downsample the image according to the perspective relationship,eliminate perspective distortion,and increase the proportion of small objects in the image.The images after perspective downsampling are input into the network model for training and testing.The mean average precision value on the test set is 99.2%,the average time is 1.89 milliseconds,and the weight is reduced from 33.8MB to 8.3MB.Experiments show that the algorithm based on perspective downsampling and improved YOLOv3-tiny can solve the problem of low accuracy,and the algorithm is suitable for deployment on low-end devices.