Research On Road Object Detection Model In Fuzzy Scene

With the continuous advancement of modern computer technology,many emerging research topics are constantly being explored.In this process,the research of target detection in the field of computer vision has always been the most popular research topic.Whether in autonomous driving or in human-computer interaction systems,target detection has played an important role,especially in recent years,the rise of various video media content and the rapid development of social networks have given people more opportunities Exposure to all kinds of images,videos and other content,make the development of the subject of the target detection field an inevitable choice.The target detection task is generally divided into two parts: one is to locate the target to be detected in the image to be detected and the accurate regression of the position,and the second is to classify the positioned target.Traditional algorithms in the field of target detection use sliding windows to select target areas for images,and then extract regional features,such as pedestrian detection models based on HOG features,and use SVM classifiers for classification.Such algorithms not only have low accuracy,but also detect Slower.After 2006,researchers slowly applied deep learning to the target detection field,such as the YOLO series detection model and the Fast RCNN series detection model.Since then,the target detection field has developed by leaps and bounds.Although the current target detection technology is very mature,whether it is the traditional detection model or the deep learning detection model,the quality of the input image of the model has a crucial impact on the detection effect of the model.For detection in some special scenarios Task,the performance of the existing model is very unsatisfactory.For example,the image with rain,fog,or night that have a lot of interference information,when using the existing deep learning domain model for detection,the detection effect often fails to reach the expected value of researcher.In order to solve this problem,this paper focuses on the detection of road targets in special scenarios,improves the existing Faster RCNN model.This paper mainly conducted the following three aspects of research work:(1)This paper proposes a feature fusion model of adaptive coefficients.By fusing feature maps at different stages in the feature extraction process to obtain feature maps with richer and more comprehensive information,the problem of excessive loss of key information during feature extraction is solved.(2)This paper cascades multiple binary classifiers in the process of selecting candidate regions that can serve higher quality candidate regions for subsequent tasks model.And make it possible to train a detection model with better performance and higher accuracy;(3)It is proposed to use local context information around the target and the target itself context information is used to assist the classification task.Combining them to solve the problems of small target detection and occlusion target detection in the detection task,and further improve the accuracy of the classification model.Finally,for the model proposed in this paper,a large number of comparative tests were conducted on the BDD100 K that is public autonomous driving data base.The experimental results prove that the model proposed in this paper can obtain better detection results in fuzzy scenarios with more interference information.