الفهرس 
Title : Hybrid model for enhancing human tracking
ContantsOnly 14 pages are availabe for public view
 |  | from | 121 |  |  |
| | | from | 121 | | |
Abstract
In this thesis, the problem of a multi-human tracking algorithm is proposed and applied to complex scenes with a single camera. Multi-human tracking mainly consists of three modules. The three modules are objects detection, data association and tracking. A new approach is proposed for a multi-human tracking problem based on a hierarchy of convolution features. We will present two approaches to show that multi-human tracking can be improved using convolution features. The first method is based on three steps. First, fast region-based convolutional neural networks is trained to detect pedestrians in each frame. Then cooperate it with correlation filter tracker which learns the target{u2019}s appearance from pretrained convolutional neural networks. Correlation filtermiddle and last convolutional layers to enhance targets localization. However, correlation filters fail in case of targets full occlusion. This leads to separated tracklets(mini-trajectories) problem. So as a third step, a post processing step is added to link separated tracklets with minimum-cost network flow. A cost function is used, that depends on motion cues in associating short tracklets. Experimental results on MOT2015 benchmark show that the proposed approach produces a comparable result against state-of-the-art approaches. It shows an increase 4.5% in multiple objects tracking accuracy. Also, mostly tracked targets is 12.9% vs 7.5% against state-of-the-art minimum-cost network flow tracker