| As a basic and important task in the field of computer vision,object detection is to judge the category of the detected object from the specified image and locate its location,which has a very broad application prospect.With the rapid development of deep learning technology,One-stage object detection algorithm based on regression analysis has developed rapidly because of its fast detection speed and high detection accuracy.Because there are quite a lot of small-scale and partially occluded objects in the complex traffic road scene,directly using the existing One-stage object detection algorithm for detection often leads to low detection accuracy and inaccurate bounding box positioning,and the use of the model will be limited by system resources.In order to improve the performance of object detection in complex traffic road scenes,this paper takes two One-stage object detection algorithms,Refine Det and YOLOv5 s,as the basic framework,and improves and evaluates the algorithm on the self-built dataset to make it meet the needs of practical engineering application,which has theoretical significance and application value.The main work and innovation of this paper are as follows:(1)Aiming at the existence of multi-scale objects in complex traffic scene,a multi-scale object detection optimization method based on improved Refine Det algorithm is proposed.Firstly,the main feature map output by ARM module in Refine Det algorithm is fused with the multi-scale feature map generated by the improved LFIP;Secondly,the ODM module in the Refine Det network structure is embedded with the improved RFB block;Then,CIOU Loss is used as the bounding box regression loss of the ODM module in the Refine Det algorithm;Finally,the algorithm is verified on the T-Monitor dataset.The mAP of this algorithm is87.2%,which is 1.1% higher than that of the original Refine Det algorithm and 3.7% higher than that of SSD algorithm.According to the experimental results,this algorithm can solve the problem of low precision of small-scale objects detection in complex traffic road scene to a certain extent.(2)Aiming at the partially occluded objects in the traffic road scene with dense objects,an optimization method of dense objects detection based on improved Refine Det algorithm is proposed.Firstly,an improved CBAM attention model is added to the TCB module of Refine Det algorithm;Secondly,the ODM module in the Refine Det algorithm model is embedded with the improved Res Block module;Then,the Repulsion Loss is used as the bounding box regression loss of ODM module;Finally,the algorithm is verified on the C-Monitor dataset.The m AP of this algorithm is 85.9%,which is 0.8% higher than that of the original Refine Det algorithm and 3.6% higher than that of SSD algorithm.According to the experimental results,this algorithm can solve the problem of low detection accuracy of partially occluded objects in the traffic road scene with dense objects to a certain extent.(3)Aiming at the low detection accuracy of lightweight object detection algorithm in vehicle driving scene,a road object detection optimization method based on improved YOLOv5 s is proposed.Firstly,the ECA module is embedded into the Bottleneck CSP1 module of the backbone network of YOLOv5 s algorithm,and a shortcut connection is added to the beginning and end of every three adjacent Bottleneck CSP1 modules to enhance the reuse of feature information in the network;Secondly,the neck network structure of YOLOv5 s algorithm is improved to the BiFPN network structure;Thirdly,the improved Aggregation Loss is used as the bounding box regression loss;Finally,the algorithm is verified on the V-Monitor dataset.The m AP of this algorithm on the V-Monitor dataset is81.5%,which is 1.3% higher than that of the original YOLOv5 s algorithm model;in terms of detection speed,the FPS of this algorithm on GTX1080 is 59.5fps,which is 12.4fps less than the original YOLOv5 s algorithm model.According to the experimental results,this algorithm can improve the detection accuracy of lightweight objects detection algorithm in vehicle driving scene to a certain extent.(4)Aiming at the problem that the object detection algorithm occupies more system resources in vehicle driving scene,a lightweight object detection optimization method based on improved YOLOv5 s is proposed.Firstly,the last three Bottleneck CSP1 modules of the backbone network of the YOLOv5 s algorithm model are replaced with corresponding Ghost Bottleneck modules,and the CBH module before the SPP module is replaced with depthwise separable convolution,and an ECA attention module is added after the SPP module;Secondly,based on the PANet network structure of the neck network,ASFF module is integrated;Thirdly,CIOU Loss is used as the bounding box regression loss of YOLOv5 s algorithm model;Finally,the algorithm is verified on the V-Monitor dataset.The m AP of this algorithm on the V-Monitor dataset is 80.5%,which is slightly improved by 0.3% compared with the m AP of the original YOLOv5 s algorithm model;in terms of the parameter size,the parameter size of this algorithm is 6.19 MB,which is 1.06 MB less than the original YOLOv5 s algorithm model.On GTX1080,the FPS of the algorithm is 74.6fps,which is 2.7fps more than the original YOLOv5 s algorithm model.The experimental results show that this algorithm can reduce the parameter size of the model to a certain extent and improve the detection speed on the premise of maintaining a certain accuracy.In general,in view of the large number of small-scale objects,partially occluded objects and the large system resources occupied by the algorithm in the complex traffic road scene,this paper proposes four improved visual object detection algorithms based on two efficient One-stage object detection algorithms,Refine Det and YOLOv5 s,and verifies the effectiveness and feasibility of the improved object detection algorithm on the self-built data set,It provides research foundation and ideas for the next research and application in traffic video intelligent monitoring,vehicle road coordination and automatic driving. |
PDF Full Text Request