Research On Video Multi-target Detection Based On Deep Learning

With the increasing popularity of urban camera installations and the development of video surveillance,video multi-target detection has become a hot research area for computers.With the development of deep learning and the perfection of computer hardware equipment,great achievements have been made in the fields of face recognition unmanned driving and target detection,and a wide range of applications have been realized.This paper focuses on the field of computer video.The deep learning algorithm is used to detect the objects in the video in real time,and the algorithm is transplanted to the embedded computing platform to complete the video multi-target detection system design on the mobile end.Considering that the parameters of the existing deep learning-based target detection algorithm are too large and the embedded computing power is limited,we pruned the algorithm according to the end-to-end prediction method of the SSD target detection algorithm,and proposed a lightweight model.The target detection algorithm MobilenetV2-SSD,and comprehensively compare the advantages and disadvantages of SSD,MobilenetV1-SSD,MobilenetV2-SSD algorithm.The MobilenetV2-SSD algorithm is implemented on the Caffe platform,and the training results are ported to the NVIDIA Jetson TX2 embedded platform to implement a video multi-target detection system on the mobile side.The main contributions of this paper are as follows:A.The target detection algorithm based on the region proposal strategy has a slow detection speed of video multi-target.In this paper,the frame is regressed by the anchor method in the SSD target detection algorithm,and the convolution is used to verify the current boundary of the convolutional neural network.Forecast and target regression strategies.B.Considering that the parameters of SSD target detection algorithm are too large,this paper proposes the MobilenetV2-SSD algorithm based on MobilenetV2's Inverted Residual Block convolutional neural network structure,which realizes the pruning and lightweight algorithm model of the model.The experiment proves that it is not obvious.When the model accuracy is reduced,it can be compressed to 1/5 of the original model.C.In order to realize video multi-target detection on the embedded side,we use the NIVIDIA Jetson TX2 embedded platform with GPU chip on the embedded side.After experimental testing,it is trained on the server and transplanted to JetsonTX2 based on the MobilenetV2-SSD algorithm.It can complete the real-time detection of video multi-target on the mobile end,and realize the purpose of video multi-target detection on the mobile end.