Research And Application Of Deep Learning Algorithms For Object Detection In Complex Scenes

With the rise of deep neural networks,object detection,as a basic task in computer vision,has developed rapidly and is now widely used in many fields such as smart cities and intelligent transportation.In recent years,intelligent unmanned equipment equipped with object detection functions has further promoted the intelligence of public security,disaster prevention and mitigation,park patrol and military applications,and has been applied to a certain extent.However,the complexity and diversity of actual scenes make object detection face huge challenges.At present,there are two main problems with object detection in complex scenes.(1)It is difficult to detect objects with fixed shapes.Take the small object in UAV scene as an example.Due to the long distance and variable angles,the size of the object is too small and its features are not obvious,resulting in a large number of missed detections in the existing algorithms.(2)It is difficult to detect objects with morphological changes.Taking flames and smoke as examples,the shapes,characteristics,and boundaries of such objects vary greatly,resulting in low accuracy of existing object detection algorithms.From the perspective of theoretical research and algorithm application,this thesis proposes three object detection algorithms in complex scenarios,and applies them to actual scenarios.The main research contents of this thesis are as follows:(1)A multi-scale small object detection algorithm based on spatial pooling Darknet53 is proposed.First,with Darknet53 as the backbone,a new spatial pyramid pooling(SPP)module has been added.SPP spatially fuses local and global features,and at the same time,it enlarges the receptive field,so that it can extract sufficient context information for small object detection.Second,a small-scale branch detector is added.This branch integrates high-level semantic information,provides more accurate location information,and solves the problem of label rewriting.Finally,the distance intersection over union(DIOU)loss is used as the bounding box regression loss of the algorithm.DIOU makes the predicted bounding box closer to the true bounding box of the object.Apply the multi-scale small object detection algorithm based on spatial pooling Darknet53 to the UAVDT(Unmanned Aerial Vehicle Detection and Tracking,UAVDT),a drone scene data set.The experimental results show that,compared with the typical algorithm YOLOv3(You Only Look Once version 3,YOLOv3)that also uses Darknet53 as the backbone,the mean average precision(mAP)of algorithm proposed in this thesis is 7.6% higher than that of YOLOv3.Therefore,the multi-scale small object detection algorithm based on spatial pooling Darknet53 proposed in this thesis not only has a good detection effect for objects on the ground,but also enhances the detection effect of small objects in the UAV high-altitude scene,which has high practical application value.(2)A multi-scale small object detection algorithm based on residual holes across the local Darknet53 is proposed.First,with cross stage partial Darknet53(CSPDarknet53)as the backbone,a residual atrous spatial pyramid pooling(RASPP)module is proposed.RASPP not only has the advantages of SPP,but also ensures that the small object details in the fusion stage are not lost.Secondly,a smaller-scale detector is added to further enhance the detection of small objects.Finally,in order to solve the problem of data sample imbalance,the focus loss function(focal loss)is introduced.Apply the multi-scale small object detection algorithm based on residual holes across the local Darknet53 to UAVDT and Vis Drone2019(Vision Meets Drone2019).The three types of UAVDT objects,the ten types of Vis Drone2019 objects,and the detection accuracy of objects in various complex environments such as day,night,and fog are calculated.The experimental results show that,compared with the highperformance algorithm YOLOv4,which also uses CSPDarknet53 as the backbone,the algorithm proposed in this thesis is better than YOLOv4 in most cases.The mAP of algorithm proposed in this thesis is 2.8%(UAVDT)and 5.05%(Vis Drone2019)higher than YOLOv4,respectively.It can be seen that the multi-scale small object detection algorithm based on residual holes across the local Darknet53 proposed in this thesis can effectively solve the detection problem of small objects in the UAV scene and has good application value.(3)A multi-scale object detection algorithm based on cross-local ResNext50 is proposed.Aiming at objects with changeable shapes,an object detection algorithm for flame and smoke is proposed.First,through a large number of video search,video framing,data screening,object labeling,etc.,a flame and smoke object data set containing 9729 pictures and 70 types of scenes was constructed.Therefore,the problem of lack of flame and smoke data sets is solved.Then,with cross stage partial ResNext50(CSPResNext50)as the backbone,combined with SPP and multi-scale detection,smooth curve Mish function,cross entropy function and complete intersection over union(CIOU)loss function are used.The whole training process adopts the strategy of Mosaic data enhancement method.Apply the multi-scale object detection algorithm based on cross-local ResNext50 to a self-made flame and smoke dataset.The experimental results show that,the mAP of the algorithm proposed in this thesis for flame and smoke detection is as high as 90.36%,which is 7.89% and 4.58%higher than that of YOLOv3 and YOLOv4,respectively.Therefore,the flame and smoke object data set constructed in this thesis solves the lack of related data sets.The multi-scale object detection algorithm based on cross-local ResNext50 proposed in this thesis can be applied to an intelligent fire detection system.In summary,this thesis proposes three object detection algorithms based on deep learning for objects in complex scenes,and apply them to actual scenes and data to obtain higher objective indicators.The multi-scale small object detection algorithm based on spatial pooling Darknet53 can be applied to detect small object in highaltitude scenes,the multi-scale small object detection algorithm based on residual holes across the local Darknet53 can be used to detect small objects in a variety of complex drone scenes,and the multi-scale object detection algorithm based on cross-local ResNext50 can detect various morphologies of flames and smoke.Therefore,the object detection algorithms proposed in this thesis have certain theoretical innovations and application values.