Research On Algorithm And Implementation Technology For Real-time Object Tracking Under Complex Environment

With the development of modern electronic information technology day by day,the applications of visual object tracking technology are growing in military, industrial and civil fields,and so on. It has been widely applied in imaging guidance, fire control systems, unmanned aerial vehicles, robots, intelligent monitoring, and so on. Though the researchers at home and abroad have been researching on tracking for decades and proposed a series of tracking algorithms, the performance of tracking is still not ideal under actual complex scenes and there are a lot of difficulties needed to be overcome, such as the target undergoes illumination variation, scale variation, pose changes, occlusion, rotation, and so on. The future development trend of tracking systems must be intelligentialize, high speed, high accuracy, universality, miniaturization, low power consumption. Thus tracking algorithms need to be better adaptive to complex scenes and more complex tracking algorithms need to be implemented in hardware systems. Therefore,the tracking technology under complex scenes has been researched and explored in this paper, including research on tracking algorithms and hardware implementation of tracking algorithms. The main research work and achievements are as follows:(1)To solve the poor tracking performance problem in traditional compressive tracking when the target undergoes appearance variation,a real-time compressive tracking method based on phase congruency is put forward. Firstly,the phase congruency transformation of the search area around the target is calculated. Then positive and negative sample sets are extracted,the samples are weighted,the positive and negative sample sets are used to train a number of weak classifiers, a certain number of optimal weak classifiers are chosen to form a strong classifier. Finally,the strong classifier is used to classify the positive and negative samples in the next frame, and therefore the target location is determined. Experimental results show that the proposed method can be well adaptive to target appearance variations,and has a good real-time performance and a high accuracy.(2)To settle the scale estimation problem in visual object tracking, an adaptive scale visual tracking algorithm is proposed in the framework of the traditional tracking with kernelized correlation filters. The tracking task is decomposed into two parts in our algorithm,including target location and scale detection. Both target location and scale detection are completed with kernelized correlation filters, and a new online update scheme is adopted. Firstly, the location and scale kernelized correlation filters are obtained by learning the regularized least squares classifiers. Then their output responses are calculated,the location and scale which make them maximum are found and they are regarded as the target location and scale. Finally, the target model and transform coefficients are updated online. Experimental results show that the algorithm is robust to target scale changes, illumination variation, pose variation, partial occlusion,rotation,fast movement and other complex scenes.(3)In order to address the poor adaptability to target appearance changes because of low-level features and easy drift when online updating in the tracking via spatial-temporal context, a spatial-temporal context tracking algorithm based on on-line detection is proposed. Firstly, a random fern classifier is built with naive bayesian classifiers, the candidate samples are classified and screened with the random fern classifier. Then the samples’ confidence is calculated by using the nearest neighbor classifier, and the confidence map of the sample whose confidence is the maximal is calculated. Finally, the obtained confidence map and the confidence map calculated from the spatial-temporal context tracker are merged into a new confidence map to determine the target. Experimental results show that the algorithm can adapt to changes in the target appearance under complex scenes, and has a good tracking performance and a good real-time capability.(4) The hardware implementation of tracking algorithms is studied and explored in this paper. Firstly, the structure of spatial-temporal context based tracking is modified according to the hardware characteristics of FPGA. Then the hardware architecture of spatial-temporal context based tracking is designed, and is implemented on FPGA. The architecture contains a preprocessing module, a priori model module, a spatial context model learning module, a target position detection module, and a tracking window overlay module, and so on. Experimental results show that the designed architecture works stably and performs well.