DEEP EXTREME TRACKER BASED ON BOOTSTRAP PARTICLE FILTER

Abstract

Visual tracking in mobile robots have to track various target objects in fast processing, but existing state-of-the-art methods only use specific image feature which only suitable for certain target objects. In this paper, we proposed new approach without depend on specific feature. By  using deep learning, we can learn essential features of many of the objects and scenes found in the real world. Furthermore, fast visual tracking can be achieved by using Extreme Learning Machine (ELM). The developed tracking algorithm is based on bootstrap particle filter. Thus the observation model of particle filter is enhanced into two steps: offline training step and online tracking step. The offline training stage is carried out by training one kind of deep learning techniques: Stacked Denoising Autoencoder (SDAE) with auxiliary image data. During the online tracking process, an additional classification layer based on ELM is added to the encoder part of the trained. Using experiments, we found (i) the specific feature  is only suitable for certain target objects (ii) the running time of the tracking algorithm can be improved by using ELM with regularization and intensity adjustment in online step, (iii) dynamic model is crucial for object tracking, especially when adjusting the diagonal covariance matrix values. Preliminary experimental results are provided. The algorithm is still restricted to track single object and will extend to track multiple object and will enhance by creating the advanced dynamic model. These are remaining for our future works.