Research On Object Detection And Behavior Analysis Method In Indoor Surveillance Scene

With the further development of intelligent surveillance technology,people pay more and more attention to indoor scene understanding and indoor safety problems,and a large number of surveillance systems are applied to the practical indoor scene such as housing and office.As the two key tasks in the surveillance system,object detection and behavior analysis has been widely studied and applied by academia and industry.In this thesis,a new object detection algorithm and two new behavior detection algorithms are studied for the practical surveillance scene,and an indoor intelligent surveillance system is designed and implemented.The main work of this thesis are as follows:1.this thesis studies a CNN object detection algorithm that combines multiple layer of convolution feature map.Firstly,the lower-level and middle-level feature map are reorganized respectively.Then,the lower-level,middle-level and high-level feature maps are connected in series,and the three layers of feature map are effectively integrated by a convolution layer.Finally,the objects is detected on the integrated feature map.Our Object detection algorithm makes full use of the low-level feature map to detect small objects and the advantages of localization,Thus can effectively improve the object detection accuracy and localization accuracy.2.In this thesis,a behavior detection algorithm based on human-object interaction is studied.To detect human hand in video is so difficult,In order to solve this problem,this thesis studies an algorithm to determine the interacting-objects based on the skin color detection results.This method can effectively detect the interacting-objects with less time.Based on the detection results of human and interacting-objects,we study a new method to modeling human-object interacting relationship based on relative spatial location.Finally,extract the feature and classify the relationship of human-object interaction,and the human behavior category is obtained.3.In this thesis,we study a human behavior detection algorithm based on CNN and the local block feature of human body.The behavior detection problem is transformed into a fine-grained object detection problem,and with the powerful feature extraction ability of CNN network,we solve the human detection and behavior recognition tasks with a single CNN network.Facing with the drawback of such detection model,we study a new method to divide the upper body of the human body,and finally use the local feature of the human body to assist the completion of thebehavior recognition task.In order to train the model and test the effectiveness of our studied algorithm,we constructs an indoor object detection dataset and an indoor behavior detection dataset respectively.In order to ensure the generalization of our algorithm,we uses the network image to construct the dataset and test it in the practical surveillance video.experimental results shows that our algorithm can effectively solve the object detection and behavior analysis tasks in different indoor surveillance scene.