**3. The kSSDAE-based tracker**

Overall structure of the proposed kSSDAE-based robust tracking algorithm for outdoor vehicle is shown in **Figure 1**. The tracking system mainly includes three parts as follows: offline

**Figure 1.** Overall structure of the proposed kSSDAE-based robust tracking algorithm for outdoor vehicle.

training of SDAE, construction of classification neural network, and estimation of object state. The basic idea of the algorithm is: firstly, we adopt the pre-trained SDAE model proposed in DLT [15] to learn the generic feature representation. Training data of the model are obtained through sampling randomly 1 million images from Tiny Images data set [26]. Tiny Images data set contains many kinds of the scene image. Before offline training, we need to preprocess the input data with 32 � 32. Offline training way of the SDAE is unsupervised. Secondly, we propose a kSSDAE model to learning more invariant feature of object appearance during tracking and train a classification neural network to compute the confidence of each particle. This is the key step to achieve robust tracking. Without the kSSDAE, the input cannot be guaranteed to have a sparse representation to extract more effective features to adapt the object appearance change. Finally, we estimate the object state under the particle filter framework, that is, the object state of the current frame can be represented by the particle with maximum confidence, which is calculated by classification neural network.

The specific implementation of the two main parts of the proposed tracking method will be stated in detail in the next section.
