Research On Target Detection Technology Based On YOLOv5

Object detection technology based on deep learning has been widely used in the field of traffic vehicle detection.This technology helps to build an intelligent real-time traffic monitoring system.However,the existing object detection models have problems such as large amount of calculation and low detection accuracy for small objects,which limit their practical application.To this end,this paper improves the design based on the existing YOLOv5-based target detection model,enhances its accuracy for small target detection,and realizes the lightweight model.Finally,the improved model is deployed to edge devices for hardware experiments verify.The main research contents are as follows:(1)Based on YOLOv5s in the YOLOv5 target detection model,to address the issue of insufficient detection accuracy of some small targets and inconspicuous targets,a convolutional block attention module(CBAM)is introduced into the backbone of the YOLOv5s model,and the feature fusion part of the model is further improved by introducing a weighted bidirectional feature pyramid network(BiFPN),it makes the model integrate more meaningful semantic information and location information during training.The improved model was trained and compared with original model on the same dataset.The results showed that the improved model had better relevant evaluation indicators than the original model,and was more accurate in identifying some small targets.The detection accuracy and generalization ability of the model were improved.(2)In order to solve the problems of high computing cost and heavy computing load in the existing target detection model,the backbone network in YOLOv5s model is replaced by the reverse residual module in ShuffleNet-V2,the integrated network is further improved in its activation function,the compressed excitation network is introduced,and the Focus module is modified to obtain a new integrated target detection model.A comparative study was conducted on the relevant models under the same experimental conditions,and the results showed that the improved model in this paper can effectively reduce the computational complexity and size of the model while hardly losing detection accuracy,achieving a lightweight and high-precision target detection model.(3)Deploy the improved lightweight model in this work to the mobile terminal,select Raspberry Pie as the mobile Edge device,install relevant software systems in Raspberry Pie to build a deep learning framework,and use OpenVINO reasoning framework and NCS2 Intel’s second-generation neural computing stick auxiliary model for reasoning acceleration.After testing,the target detection results of Raspberry Pi end showed good detection performance and real-time performance.