Human Detection And Tracking Based On Visual Saliency

With the increasingly development of intelligent technology,service robots are gradually entering people's daily life.More intelligent for robots is required to meet the needs of human being.Similar to vision system of human being,vision sensor is the main perception way of external environment information.Since robots need to recognize objects in an environment when they provide services,vison-based object detection is an important topic in the field of robotics.In addition,as a key technology to realize the interaction between robots and humans,human tracking has attracted more and more attentions.However,how to accurately and quickly detect and track the targets in the scene over a large amount of data becomes a chanllege.Human visual system has a powerful ability of data screening,so that human beings can quickly and accurately find the main information in the scene and ignore other non-important information.This ability is called visual selective attention mechanism or visual saliency mechanism.In this thesis,we investigate human detection and tracking approaches based on human visual saliency mechanism,and explore the application in mobile robots.The main contents of this thesis are as follows:(1)To rapidly localize objects for robots,a novel saliency detection method based on global multi-scale surpixel contrast is proposed.First,considering that the saliency detection method based on regional characteristics is based on cognitive psychology,which achieves the localization of sailent region by introducing competition mechanism,and the objects in the scene generally appear in different scales,we segment input images using simple linear iterative clustering algorithm with different parameters.Then,considering that the histogram of color feature is one of the most effective visual saliency descriptors,we extract the color histogram feature of each region in different scales and compute saliency value of each region using spatial information and the prior position.Finally,in order to make use of the advantages of saliency maps in different scales,a fusion method based on multi-layer cellular automata is applied.The experimental results on several datasets show that the proposed method obtains better performance compared with other popular methods.(2)To make the robot recognize human body accurately and quickly,a novel humam detection method based on visual saliency and multi-feature fusion is proposed.First,locate the regions of interest and ignore background information using the saliency detection method.Then,process the regions of interest and construct human detection model.In consideration of the diversity of human body postures and appearance,we propose a multi-feature fusion method to model human beings.Finally,the histogram of oriented gradient feature and color feature and the second order gradient feature based on cell are extracted from the positive and negative samples and the features are selected using the boosting method,and the SVM is also trained to perform human detection.Experimental results on several datasets show that the proposed method achieves higher detection accuracy and efficiency compared with other popular approaches.(3)To enable the robot to track automatically and effectively deal with the mutation movement in tracking process,we propose an adaptive kernelized correlation filter tracking method based on visual saliency.First,we use visual saliency detection to initialize achieve target instead of manual calibration.Then,a human body is described by using multi-features.Next,human body tracking is performed by using mutil-channel kernelized correlation filter.In the tracking process,the visual saliency detection method is used to relocate and track the target when drift occurs.The experimental results on different image sequences show that the proposed method can automatically relocate and track the target,when the human body occurs mutation and even the target is lost.(4)A demo software system is designed and developed to effectively valuate the proposed methods mentioned above by using MATLAB GUI technology.The software system mainly consists of five parts:data selection and display module,saliency detection module,human detection module,human tracking module and control module.By simple and intuitive operations,object detection and tracking in different datasets and video sequences are performed.By using the image sequences captured by the Pioneer 3-DX robot,salient objects detection and human tracking in the real scene are realized.