Research On Object Detection Methods In Traffic Scenes Based On Deep Learning

Automatically detecting various objects(such as vehicles and pedestrians)in images or videos from traffic scenes is an essential prerequisite for many intelligent transportation systems.Reasonable traffic management and control based on the movement of vehicles and pedestrians can reduce the occurrence of incidents such as road congestion and traffic accidents and protect people's lives and property.With the rapid development of computer technology and the extensive application of computer vision theory,using machine learning and deep learning to classify,locate or track objects in different scenes has become a hot research topic in recent years,and real-time object detection has important application value in self-driving,intelligent video surveillance,industrial detection,military object detection and so on.Although deep learning has brought revolutionary progress to the field of object detection,there are still many shortcomings in the existing algorithms based on deep learning,such as the computational complexity of the algorithm is too high to be applied in real-time,or the detection effect of small objects is poor,that is to say,these methods are difficult to achieve the trade-off between detection accuracy and speed.In order to solve these problems,this thesis uses the idea of feature pyramid to detect all kinds of objects by using the information extracted from multi-layer feature images,which can effectively reduce the occurrence of missed detection and false detection.Specifically,the research contents of this thesis mainly include the following points:Firstly,aiming at the shortage of dataset in traffic scenes,this thesis constructs an urban traffic dataset using traffic surveillance videos,dividing the training set and the testing set,and uses the annotation tool BBox-Label-Tool to classify and annotate the images in the dataset,then these images can be used for the subsequent network training and testing.Secondly,because the scale of different objects varies greatly in traffic scenes,a 32-layer multi-branch object detection network is proposed in this thesis to detect objects of different scales.Through the reasonable design of three detection branches,the model can accurately detect large,medium and small scale objects in different scenes such as sparse,crowded,daytime and nighttime.Compared with other methods,this method improves the detection precision and speed,which realizes a good balance between them.Finally,in view of the problem that the above method does not make full use of the contextual information,the detection results of some objects(especially small objects)are not accurate enough by this method,in this thesis a multi-feature fusion based object detection method is proposed to introduce the contextual information through the fusion module,which significantly improves the model's detection effect of various objects,especially small objects.The experimental results on different datasets show that the performance of this method on different evaluation metrics has been improved accordingly compared with other methods.