**4.1. Quantitative evaluation**

In this chapter, we adopt two quantitative evaluation indicators: one-pass evaluation (OPE) of tracking precision and success rate [28]. The precision takes the position error as the benchmark, and the precision plot shows the percentage of frames whose estimation position error is less than the given threshold, and the horizontal axis of the precision plot is scaled to the range [0,50]. The success rate is based on the overlap rate, and the success plot counts the number of successful frames whose overlap is greater than the given threshold, and the horizontal axis of the success rate is scaled to the range [0,1]. We use the score for the threshold = 20 pixels of each precision plot and the area under curve (AUC) of each success plot to rank trackers and onepass evaluation (OPE) for robustness evaluation. The scores and rankings of precision and success rate for four trackers on the overall performance and the 11 attributes performance are shown in **Table 1**, and the best tracking results corresponding to the overall performance and the 11 attributes are marked in bold, and the ranking score is shown after "\." In **Table 1**, SV: scale variation, OV: out-of-view, OPR: out-of-plane rotation, OCC: occlusion, LR: low resolution, IPR: in -plane rotation, IV: illumination variation, DEF: deformation, MB: motion blur, BC: background clutters, FM: fast motion. The precision plot and success rate plot of four trackers on overall performance is shown in **Figure 3**. The precision plots and success plots of


**Table 1.** Tracking performance on four trackers.

**Figure 3.** The overall performance of precision plot and success rate plot on four trackers.

four trackers on 11 attributes performance are shown in **Figure 4**. In order to analyze the performance of the tracker in every challenging attribute, [28] has marked the characteristics of each sequence and constructed subsets of the sequences with different saliency characteristics. For example, the OCC subset includes 29 sequences; it can be used for analyzing the ability of the tracker to handle occlusion problem. In **Figure 4**, the number that appears in the legend of each graph represents the ordinal number of sequence subset.

In overall performance of precision and success rate, our tracker is significant higher than the other three trackers. The performance of our tracker ranks first in 8 out of 11 attributes on precision as shown in **Table 1**. At the same time, the performance of our tracker ranks first in 6 out of 11 attributes on success rate. For the other attributes, except for LR attribute, our tracker has the success rate very close to the best on SV and BC attributes. The success rate of our tracker ranks 3 on MB attribute, but it is only lower than the best (CSK), 2.7%. Therefore, it can be concluded that the proposed kSSDAE-T tracker is the best compared with DLT, CSK, and MTT.

According to the precision plot and success rate plot of the four properties based on OCC, IV, MB, and FM attributes, we can see that our tracker can handle the appearance changes caused by most of outdoor environmental factors.

All in all, the proposed tracker can learn the invariable features of the object appearance and deal with the problem of object appearance changing caused by most of the outdoor complex environments. It can achieve better tracking results under most outdoor challenging conditions.

#### **4.2. Qualitative evaluation**

In order to further verify the effectiveness of the proposed tracking method in real scenarios, we compared the four trackers (proposed kSSDAE-based tracker, DLT, CSK, and MTT) on four outdoor vehicle sequences in real scenarios (Car4, CarDark, CarScale, and Suv). The attributes

**Figure 4.** The precision plots and success plots of four trackers on 11 attributes performance (SV, OV, OPR, OCC, LR, IPR, IV, DEF, MB, BC, and FM).


**Table 2.** Attributes of four sequences.

**Figure 5.** The sampled tracking results. Frame numbers are shown in the top left of each figure.

of four sequences are listed in **Table 2**. The partial tracking results of the four video sequences are shown in **Figure 5**.

In the video sequence Car4, when emerging IV and OCC near the #186, #233, and #318 frames, it can be seen from the tracking results that the CSK and the MTT tracker have different degrees of object drift. But, our tracker and DLT have achieved better tracking results. In addition, our tracker can also accurately track the target vehicle when SV emerges in #321, #612, and #635 frames. In the video sequence CarDark, our tracker can still perform effective tracking when the IV and BC emerging in #62, #152, #278, #301, #387, and #392 frames, while the MTT and CSK trackers have track drift at #301 frame, and at #387 frame, they completely lost the target vehicle. In the video sequence CarScale, our tracker can still show great performance when OCC was occurred in #165 and #175 frames, but CSK tracker failed. In the video sequence Suv, despite the OCC and similar background interference, our tracker can still accurately track the target vehicle.

To summarize, the proposed kSSDAE-based tracker can perform well in most complex outdoor environment.
