Research And Improvement Of Compressive Tracking Algorithm

Visual tracking technology has been the research hotspot and difficulty in the field of computer vision, which has been extensively studied in recent decades because of its importance in practical applications such as video surveillance, human computer interaction, visual navigation,traffic monitoring. Despite extensive research in this topic with demonstrate success, it has already exist some challenging factors such as illumination variation, occlusion, background clutter and shape deformation. All those factors affect the performance of tracking algorithm. Therefore, how to design an effective tracking algorithm still has important theoretical significance and practical significance. The main works of this thesis are as follows:Firstly, we have studied the theory and status of visual tracking algorithm, the relevant theoretical knowledge of the object tracking algorithm are introduced, since the proposed algorithm based on particle filter framework, the particle filter object tracking algorithm is also introduced in detail.Secondly, tracking-by-detection methods have been widely studied and some promising results have been achieved. These methods use discriminative appearance models to train and update online classifiers. They also use a sliding window to detect samples which will then be classified. Then, the location of the sample with the maximum classifier response will be selected as the new location. Compressive tracking was recently proposed with an appearance model based on features extracted in the compressed domain. We use a very sparse measurement matrix that satisfies the restricted isometry property (RIP) in compressive sensing theory, thereby facilitating efficient projection from the image feature space to a low-dimensional compressed subspace. Tracking formulated as a binary classifier problem, the positive and negative samples are projected with the same sparse measurement matrix and discriminated by a simple naive Bayes classifier learned online.Finally, as compressive tracking uses a fixed-size window to extract object feature may cause tracking drift problem. In this paper, we proposed a scale-adaptive tracking algorithm which combines the random projection-based appearance model with the particle filter framework. The proposed scale-invariant normalized rectangle feature was adopted to characterize a target object of different scales to be tracked. This represents for scale variation or in-variation targets. The compressive features of a particle in special scale were achieved by the projection via a modified adaptive random measurement matrix. This matrix is updated based on the initial matrix and the corresponding particle’s current scale value. A 2-order transition model which considers two previous statuses has been used to estimate the current position and scale status for each particle. This introduces the velocity information of the moving target into the sampling process of candidate samples both in terms of position and scale. We used an observation model based on the naive Bayesian classifier response to acquire particle importance weights and to re-sample the particles. This allowed us to preserve the important samples and relieve the drift. The experimental results showed that the proposed algorithm preserves the real-time property and performs favorably against several state-of-the-art tracking algorithms on challenging sequences in terms of accuracy, stability and robustness.