Research And Implementation Of Parking Management Vision Algorithm In Complex Scenes Based On Deep Learning

With the rapid development of the social and economy of our country,the number of motor vehicles has increased,and the demand for parking has also increased rapidly.However,there are still many limitations in the current parking management system,such as insufficient intelligence,low efficiency,and insufficient applicable parking environment.For example,the license plate recognition that is currently widely used.First of all,it is difficult to detect in large angles and complex scenes.Second,it is only applicable to parking garage entrances and exits with fixed locations and standard equipment,it cannot be applied to the detection of on-street parking.At the moment,deep learning is developing rapidly,it has played an important role in many industries and has made significant progress.It also has applications in the field of parking management.At present,vehicle detection mainly uses object detection algorithm.At large angles,there are many overlaps between the bounding boxes of the vehicle,and the features in the boxes cannot accurately represent the vehicle’s features.At present,the commonly used object detection algorithms mainly extract features in the bounding box of the target,which is not suitable for vehicle recognition at large angles.In order to solve the problem of large-angle vehicle detection,a new object detection algorithm is designed in this paper.The following are the main contents of this thesis:1.This article analyzes the basic ideas and algorithm characteristics of several classic object detection algorithms in detail,and reviews the development of object detection algorithms.Several common object detection algorithms are introduced,including R-CNN series,YOLO series,Corner Net,Center Net,etc.,and analyzed the advantages and disadvantages of each algorithm.2.Aiming at the problem of vehicle detection in large-angle and complicated situations,we analyze the shortcomings of existing target detection algorithms.At large angles,the features in the vehicle’s bounding box are more complex,including not only some other vehicle features,but also a lot of background features.The Anchor-based object detection algorithm mainly extracts the image features in the object bounding box.The object features learned by this method are not suitable for large-angle vehicle detection;the Center Net algorithm in the Anchor-free class focuses on the object center point feature and is suitable for large-angle vehicles detection,but the prediction of the width and height of the bounding box is not accurate enough,and there is room for improvement.Both types of target detection algorithms have deficiencies,this thesis combining the characteristics of the large-angle vehicle detection problem,designs a new deep learning-based object detection algorithm LAVD(Large Angle Vehicle Detection)based on the classic algorithms YOLO and Center Net.The algorithm is based on the detection of the center point of the target by the Center Net algorithm,which makes the algorithm pay more attention to the feature extraction of the central area,and is suitable for the large overlapping area of the bounding box of the large-angle vehicle.Second,it adds the anchor idea in the YOLO algorithm to To solve the problem of missed detection of overlapping targets inherent in Anchor-free algorithms,and in order to detect more accurate width and height values,this paper uses anchor boxes to detect the width and height offset values of the target bounding box,which is different from the detection method that Center Net directly obtains the width and height values through neural network calculations.3.In this thesis,a corresponding data set is produced for vehicle detection in largeangle and complex scenes.The data set contains parking images under different angles,different backgrounds,and different degrees of occlusion,a total of 1617 images.And the YOLOv3 and Center Net algorithm models are reproduced on this dataset and the object detection standard dataset Pascal VOC,and the experiment is compared with the large-angle vehicle detection algorithm proposed in this paper.Experimental results show that our method has a better effect on the detection of large-angle vehicles,and also has a better detection accuracy rate on the standard data set.The network can realize real-time largeangle vehicle detection and has certain generalization ability.