Author Affiliations
1College of Software, Liaoning Technical University, Huludao, Liaoning 125105, China2Beijing Chaoxing Company, Beijing 100000, Chinashow less
Fig. 1. Flow chart of our algorithm
Fig. 2. Schematic diagram of the context overlap sampling
Fig. 3. Tracking instance and its corresponding response graph. (a) Normal tracking; (b) occlusion occurs; (c) response graph when normal tracking; (d) response graph when occlusion occurs
Fig. 4. Comparison and verification of auxiliary module and feature selection. (a) DP; (b) OPE
Fig. 5. Tracking results of different algorithms on the OTB2015 data set. (a) DP; (b) OPE
Fig. 6. OPE of different algorithms in different scenarios. (a) BC; (b) DEF; (c) OCC; (d) SV; (e) OPR; (f) OV
Fig. 7. Tracking results of different algorithms on 6 video sequences. (a) Biker; (b) Bird1; (c) ClifBar; (d) Ironman; (e) Soccer; (f) Swinning
Initialization parameter | Value |
---|
λ | 0.01 | λ1 | 0.3 | λ2 | 0.03 | ρ | 0.5 | β | 0.075 |
|
Table 1. Initialization parameters of the main module
Initialization parameter | Value |
---|
λ0 | 0.01 | η | 0.075 | Kernel function | 0.5 |
|
Table 2. Initialization parameters of auxiliary module
λ2 | DP/% | AUC/% |
---|
0.001 | 76.1 | 55.6 | 0.01 | 79.9 | 59.5 | 0.03 | 83.7 | 61.4 | 0.05 | 81.0 | 59.8 | 0.07 | 79.8 | 59.8 | 0.10 | 79.0 | 58.5 |
|
Table 3. Results of the tuning experiment
Module | Feature | DP/% | AUC/% | FPS/frame |
---|
CACF | HOG+CN | 76.1 | 58.3 | 54.0 | Main module | HOG+CN | 83.7 | 61.4 | 48.3 | Auxiliary module | Conv4-3 | 90.4 | 65.1 | 2.7 | Main+auxiliary | HOG+CN,Conv4-3 | 87.2 | 63.7 | 17.0 |
|
Table 4. Independence verification results of modules
Algorithm | DP/% | AUC/% | FPS/frame |
---|
Staple | 78.4 | 57.9 | 57.3 | MEEM | 78.1 | 53.0 | 16.8 | DSST | 68.7 | 51.7 | 20.1 | SRDCF | 78.8 | 59.8 | 11.5 | KCF | 69.5 | 47.7 | 114.0 | Ours | 87.2 | 63.7 | 17.5 |
|
Table 5. Comparison between our algorithm and the tracking algorithms with better real-time performance
Algorithm | DP/% | AUC/% | FPS/frame |
---|
LSART | 92.3 | 67.2 | 0.1 | ECO | 90.9 | 68.7 | 1.4 | CCOT | 89.6 | 66.7 | 0.3 | Ours | 87.2 | 63.7 | 17.0 |
|
Table 6. Comparison between our algorithm and the algorithms based on deep convolutional neural network
Tracker | EBT | CCOT | Staple | Struck | CSR-DCF | SRDCF | KCF | Ours |
---|
EAO | 0.291 | 0.331 | 0.295 | 0.142 | 0.338 | 0.247 | 0.192 | 0.327 | A | 0.441 | 0.539 | 0.545 | 0.439 | 0.510 | 0.535 | 0.491 | 0.522 | R | 0.920 | 0.238 | 1.350 | 3.370 | 0.850 | 1.500 | 2.010 | 0.814 |
|
Table 7. Tracking effects of different algorithms on the VOT2016 data set