Research On Sample Vehicle Detection Technology For Natural Scenes

In recent years,with the rapid development of computer technology,more and more image processing technologies have been added to the intelligent identification system.The identification system can increase the monitoring range while improving the monitoring accuracy.Vehicle detection is an important branch in target detection.The automatic detection and identification of vehicles plays an important role in military and civilian applications.In military applications,it can be applied to target detection,tracking,detection and guidance,etc.,real-time monitoring and tracking of targets,and real-time precision strike capability of weapon systems in different environments.Civilian aspects can be applied to areas such as driverless cars and urban intelligent transportation.The existing traditional detection methods detect targets under a small number of sample conditions,and have problems such as high accuracy,low recall rate,and low detection accuracy and high recall rate.The method of detecting deep learning requires a large number of samples but a large number of samples cannot be obtained in the military field and can only be detected under small sample conditions.In view of the above problems,this paper proposes a small sample detection method based on natural scenes.The main work and innovations are as follows:1.The traditional detection method detects that the Haar feature detection vehicle has a high accuracy,a low recall rate,and a low accuracy and a high recall rate.This paper proposes to use the support vector machine to classify the Haar feature detection results,remove the false alarm and detect the vehicle,and improve the detection accuracy when the detection result keeps the recall rate high.After image preprocessing,the accuracy of Haar-based feature detection method can be improved.2.Deep learning method detection,the direct detection of SSD(Single Shot MultiBox Detector)network on multi-layer features is improved to detect on the feature pyramid network structure.The improved network structure combines the features extracted by multiple scales,and the network can extract richer fine-grained feature information.The improved network structure has an improvement in detection accuracy.3.Transfer learning mode detection,using the pre-training weights on the ImageNet dataset,using the residual network(Resnet34)as the feature extraction network instead of VGG16 in the improved SSD network.The experimental results show that the network detection speed is improved without the network accuracy falling.