Automatic Detection And Tracking In The Cluttered

We research on automatic detection and tracking in the cluttered to investigate thefundamental principles and mechanisms in detection and tracking. Our ultimate aim is tofunction the computer like human eyes which can fulfill the stable and valid detection ofcommon objects. Traditional feature extraction and tracking methods under target framework arereviewed first. Traditional methods often fail under conditions that there are changes in gestures,illumination or occlusions of targets. The tracking failure is due to the bad invariance of featureextraction as analyzed in the thesis. We propose a novel tracking method which succeeds intracking without any prior knowledge, thus making the method generally applicable.We use the well-known TLD method as an example and carry on detailed tests on it. Thereare two main shortcomings in the TLD method. It is easily affected by initializations; on theother hand, TLD method is encountered with losing targets when targets rotate or changinggestures. As to the first problem, we propose an improved tracking method based on salientdetection for preprocessing. The scenes are divided into simple scenes and complex scenes.Automatic tracking is applied on simple scenes, while complex scenes need human revise. Wecome up with another improved method which is combined with salient detection andcomparisons of histograms’ similarity to solve the second shortcoming. Experimental resultsshow an improved performance. Nevertheless, the contradiction between the discrimination andthe extensibility of the features of the target remains unsolved. We further put forward an ideathat combines HMAX based detection model with TLD. According to our speculation, a jointwith features and feature libraries can reach the stable target tracking. However, the realizationof our idea still faces some problems. First, the feature library which needs a huge amount ofdata is hard to build. Moreover, the biological presentation of the brain model is severelydegraded by SVM in HMAX. Not to mention that any methods based on TLD framework couldnot essentially improve tracking results.A novel method based on biological vision is proposed in this paper to solve the problemsmentioned above. We make use of salient detection which extracts features of objects like color,texture, intensity.etc to obtain the location and a rough outline of the object. A static salient mapand a dynamic salient map are subsequently built. The static salient map is composed of salientmaps of color, orientation and intensity, while the dynamic salient map utilizes motioninformation and continuity within video frames. Then we integrate the salient maps by HFT(Hypercomplex Fourier Transform) to form the location map which presents the probability ofthe object staying at one location during the motion. The biologically based model is finally built by using the persistence of vision and redetection model.We carry out an overall test on the new method with tremendous video datasets comparedwith TLD. Our new method is proved be superior in conditions that different objects (rigid orflexible) exist, rotation of objects and long obscuration of objects,.etc. The method, however,has a key assumption that the object is salient. Thus, we summarize conditions when our methodis applicable. The new method proposed in this paper casts insight into areas that traditionaltracking methods in computer vision can be substantially improved with biological vision.Whether related research of neuroscience and brain science can be applied in tracking needsfurther investigation.