**5. Conclusion**

In this chapter, we propose an improved auto-encoder–based approach for robust outdoor vehicle visual tracking. Our tracker can adapt the change of the object appearance during the tracking. The quantitative analysis on a standard evaluation platform shows that our tracker has a better tracking performance compared with the other three state-of-the-art trackers and has higher tracking precision in most of the outdoor vehicle tracking challenges. The qualitative analysis on four outdoor vehicle sequences in real scenarios shows that our tracker can work well in most complex outdoor environment.

The unsupervised training of kSSDAE requires that bottom image cannot be too large, otherwise it will consume a lot of training time. The training data are obtained by down-sampling directly from a full-sized image leading to information loss. In order to avoid loss of input image information, we can further improve the performance of outdoor vehicle tracking algorithms by using stacked convolutional auto-encoder (SCAE) [31] to take the outdoor vehicle tracking algorithm into application of life and industry.

Note: this chapter is an extended version of [32].
