Visual Tracking Based On Deep Detection Architecture

Visual tracking is one of the fundamental problems in the field of computer vision and has a variety of applications,such as human-computer interactions,smart surveil-lance systems and autonomous driving.Due to uncertainty of the target and diversity of the scene in tracking task,it remains a challenging problem on how to keep the accuracy,real-time and robustness of the tracker under different challenges.This dissertation,considering the practical situation,aims to build a tracker which balances robustness and real-time performance.With the deep research of some difficult problems in visual tracking,a few results have been achieved for a value of applications.The main work and academic contributions can be summarized as follows:1.We propose a tracker based on object detection architecture Single-Shot Multi-Box Detector(SSD)for target feature extraction and location by applying similarity learning in visual tracking.This tracker uses three loss function for joint learning in optimizing strategy,which overcomes the disadvantages of heavy computation and hard training in similarity learning,and effectively accelerates convergence speed of network and improves performance of the tracker.Experimental results prove the effectiveness and feasibility of the tracker.2.A non-local mean based Discriminative Correlation Filter(DCF)tracker is pro-posed for visual tracking.This tracker uses a fully-convolutional DCF layer,which is composed of two separate convolution sequences,for capturing(target's)local response map and a non-local layer for(background's)global response map.High-resolution finial response map is an additive combination of local and global response map,which improves the ability of the tracker to distinguish the target from the background.The results on the benchmarks show that this tracker gains much improvement in accuracy and robustness when compared with the baseline.3.A feature pyramid based method is presented for size estimation.This method replaces image pyramid with feature pyramid in the process of target size estimation,saving much redundant computation in feature extraction caused by image pyramid and speed up tracking efficiency.Also,applying more dense feature pyramid in feature extraction for better performance will not significantly increases computation.Experi-mental results show that this tracker spends less time in the process of size estimation and keeps the same accuracy after feature pyramid is applied.