Study On Surveillance Video Object Detection Based On Transfer Learning

Video surveillance as an important part of the security field,it's development is increasingly mature and intelligent.Intelligent video surveillance analyzes the eontent of surveillance video through artificial intelligence technology,and realizes automatic identification and positioning of object,object detection has become one of hottest research points.This paper adopts the method of transfer learning and takes the surveillance video on campus as the research object.Aiming at the surveillance video in the daytime and at night,different improvement strategies are proposed from the network structure on the basis of SSD to improve the effect of object detection.The main research contents of this paper are as follows:(1)The classical object detection algorithm based on convolution neural network and the application of SSD in object detection are deeply analyzed,and different methods of transfer learning are studied,which lays a foundation for the improvement of SSD.(2)For surveillance video during the day,aiming at the problem that SSD network lacks the ability to detect small and medium object,an improved SSD multi-scale detection method based on parameter transfer is proposed.For the lower convolution layers,on the basis of preserving the lower features used for detection in SSD,the regional feature magnification extraction method is further used to enrich the detailed information of features to improve the detection effect of small object.For the higher convolution layer,multilayer feature extraction is used to increase the SSD network depth and enhance the feature extraction ability of SSD network for medium object.Finally,the improved lower and higher feature layers are combined to reconstruct the multis-cale detection convolution layers to improve the ability of object detection.(3)The domain adaptation approach is adopted in this paper,and SSD network is improved to improve its adaptability between different domain,and the surveillance video in the daytime and at night are detected at the same time.The day and night data sets are input into SSD,and MMD is applied to the specific convolution layer to reduce the differences in learning feature between different domain,this difference is incorporated into the total loss function of the network,and the loss function value is reduced by adjusting parameters through multiple iterations to get the final model.The experiment verifies the feasibility and effectiveness of the improved network.