**4. Experiments**

In this section, we conducted a quantitative experiment to evaluate the proposed tracker (kSSDAE-T) on a popular single-object online tracking benchmark [28]. The benchmark data set provides 51 fully annotated video sequences that have the 11 challenging attributes. Most of these attributes exist in the real scene of outdoor vehicles. In order to better demonstrate the performance of our tracker, we compare our tracker with other three popular trackers, including deep learning tracker (DLT) [15], multi-task tracker (MTT) [29], and Circulant Structure of Tracking-by-Detection with Kernels (CSK) [30].

The main related parameters in our experiment are set as follows.

• Learning rate is set to 0.2; sparsity k is set to 40.

