Research Of Object Tracking Method Based On The Information Entropy Features Fusion And KNN-SVM

Traditional video monitoring systems involve intensive manpower and lots ofmaterial resources, since they need a large number of staff to monitor the videosequences collected by cameras. With the development of computer vision, artificialintelligence and other related research fields, the intelligent monitoring has beendeveloped. Literally speaking, intelligent monitoring refers to the automatic analysisand recognition in the video sequences collected by cameras; the effect of identifyingand predicting the abnormal situation in the scene. Furthermore, it can track a target’sposition and predict the target trajectory. This can not only liberate the eyes of videomonitoring personnel fundamentally, but also improve the efficiency use of camera.Among may components in intelligent monitoring, object tracking is one of the mostimportant parts, which can accurately track one or more objects in the scene and thenpredict the object trajectory. Object tracking algorithms play an important role inintelligent monitoring, because it can be seen as the brain of cameras, which canupdate the parameters generated in the process of tracking to ensure the real-time andthe accurate tracking of the object by recording and analyzing.Nowadays, as study on object tracking algorithm is in full swing, a large numberof excellent object tracking algorithms have been proposed. Object tracking algorithmcan be divided into several categories from the algorithm’s perspective:(1) objecttracking based on model. The model is established through certain prior knowledgeand the atoms in the model renew unceasingly along with the tracking process;(2)object tracking based on the characteristics. Most of this kind of object trackingusually adopt the bottom-up to reveal characteristics which can describe the objectdetails more clearly, such as Haar feature and HOG feature;(3) object tracking basedon the area. Through the effective modeling of the object and surroundingbackground, trace templates can be obtained. Then we can match it in the subsequenttracking process;(4) object tracking based on the contour. The change of the object contour in the tracking process can be obtained through the adaptive change curve.The second tracking method is used in this thesis which adopts multi-featuresfusion to express objects. This can effectively avoid the shade appearing in thetracking, the deformation of the object and the change of light. The sparse image canobtain the important information of the object interested with less data. But during thetracking, it is inevitable to loss some information in the process of obtaining theimportant information because of the interference of the background content. As aresult, the object information weakens in the process of sparse. And then, these comesthe tracking error accordingly. The visual salience map can extract significant area inthe in the picture pertinently, but the object in the scene is not the most significant inthe tracking process because of the interference from other objects and background inthe scene. To achieve the fusion of both kinds of information, in this thesis wemeasure the important extent of these two characteristics above in the image trackingin information entropy. When the object is the most significant, the significantinformation of picture can be used to make up for the loss of object information in thesparse representation; when the object is not the most significant, the importantinformation extracted by the sparse representation can reduce the error brought by thesignificance of background again. KNN(K-Nearest-Neighbor) is the most classicalclassification algorithm for its quick classification of the samples, but itsdisadvantage is that the classification precision is not high; although the classificationprecision of SVM is high, nevertheless the speed is slower. So there comes theeffective combination of the two: KNN-SVM classifier, which not only takes theadvantage of both, but also makes up for the disadvantage of both. Informationentropy is adopted to make the fusion between sparse coe fficient and visual saliencefeatures. Then the result is combined with the object tracking algorithm KNN-SVM.Next, through the adaptive combination of sparse coefficient features and visualsalience features (information entropy), issues like change of light, occlusion andcomplex background in the object tracking process can be well handled. In addition,this conclusion can be proven by a large number of experiments. This method whichcombines sparse representation with visual salience in the tracking ca n make up forthe defects of both and improve the final tracking results greatly. The targets of object tracking can be face, vehicle, pedestrian and so on. A greatquantity of experiments confirmed that the algorithm in this thesis could be used inface tracking, so as to be applied in face attendance system and mobile app lications;the algorithm can also track the rigid objects, therefore we can apply it in intelligenttransportation system and recording the emergency in the driving effectively, whichcan help the police track and investigate the unusual vehicles. The computer visiontheory plays a great role through object tracking. The analysis and identification ofthe objects are easily achievable through the record of camera device and thehandling of tracking algorithm. This process can preferablely embed the objecttracking into our mobile terminals, and this can finally make it realizable that utilizingthe program to keep tracking at any time and any place.