*4.2. OTB Dataset Experiments*

In this paper, experiments are conducted on the popular OTB-50 and OTB-100 datasets in the field of target tracking, which consist of 50 and 100 fully annotated videos, respectively. In this paper, the accuracy maps in one-pass evaluation (OPE) are used to evaluate different trackers and are compared with 10 advanced trackers SiamFC, attention-based trackers MemTrack [21] and MemDTC [22], correlation filter-based trackers KCF [23], Staple [24], DSST [25] and SRDCF [26], deep learning and correlation filter-based tracker CF2 [27], CREST [28], and CSR-DCF [29] were compared for the results. As shown in Figures 6 and 7, the performance of the proposed tracker (Ours1) in this chapter is at the advanced level in both benchmark tests. Specifically, the proposed algorithm obtained success rate scores of 0.655 and 0.643 on OTB-50 and OTB-100, respectively, and the proposed algorithm gained 4.6% and 6.0% improvement over the Siamese network-based tracking method SiamFC, which confirms the advantages of the lightweight target-aware attention learning network and attention learning loss function proposed in this paper. CF2 algorithm uses the depth features of three layers in the VGG-16 network for target modeling to improve the discriminative power of the model, and obtains success rate scores of 0.603 and 0.562 for OTB-50 and OTB-100, respectively, and the performance of the proposed algorithm in this paper is 5.2% and 8.1% higher than that of the CF2 algorithm without using more depth features. The CREST algorithm achieves a higher success rate than the CF2 algorithm on the OTB-50 dataset and performs better than the algorithm proposed in this paper in terms of both success rate and accuracy; the reason for this is that the CREST algorithm introduces a residual network to extract the depth features of the target, and the residual network structure can be used to build a deeper network to

improve the accuracy of the features and alleviate the gradient disappearance problem caused by the deep network.

**Figure 6.** Success and precision rates on the OTB50 dataset.

**Figure 7.** Success and precision rates on the OTB100 dataset.

For object-tracking algorithms, the real-time performance should also be used as one of the criteria for evaluating tracker performance. In Table 1, we compared the operational performance of some of the advanced trackers in terms of Precision score (%), Success rate (%), and Speed (FPS) on the OTB-100 dataset. Table 1 shows the results of our tracker compared with 7 advanced trackers including BaSiamIoU [30], ATOM [31], CFML [32], CREST [28], CSR-DCF [29], SRDCF [26], and SiamFC [9]. From Table 1, we can note that ATOM draws on the IoU-Net idea and proposes IoU modulation and IoU predictor to solve the scale challenge in the tracking process, achieving better tracking performance in terms of Precision score and Success rate. However, the speed performance of ATOM is not as satisfactory as our tracker. Meanwhile, although SiamFC is capable of reaching 102.3 FPS in speed, it is not able to adapt to changes in target appearance during tracking, resulting in lower tracking accuracy. Our tracker achieves 83.3% in Precision score and 64.3% in Success rate in 59 FPS. Overall, our tracker strikes a balance between Precision score, Success rate, and Speed. Therefore, for some scenes with higher requirements on tracking speed, SiamFC algorithm is a better choice, while for some scenarios where tracking accuracy is more preferred, ATOM algorithm should be chosen. Our method is more suitable for applications that require a certain degree of tracking accuracy and tracking speed.


**Table 1.** The real-time performance of the advanced trackers on the OTB-100 dataset. In the table, red, green and blue indicate the top three scores respectively.
