Author Affiliations
1School of Communications and Information Engineering, Xi'an University of Posts and Telecommunications, Xi'an, Shaanxi 710121, China2School of Physics and Electronic Information, Yan'an University, Yan'an, Shaanxi 716000, China3School of Computing and Engineering, University of Huddersfield, Huddersfield HD1 3DH UKshow less
Fig. 1. Scale change of target motion in panoramic video
Fig. 2. Panoramic image stitched by seven cameras
Fig. 3. Visualization results of deep features in convolutional neural networks. (a) Conv 1; (b) Relu 1; (c) Pool 1; (d) Conv 3; (e) Relu 3; (f) Pool 3; (g) Conv 5; (h) Relu 5; (i) original image
Fig. 4. Network architecture of proposed algorithm
Fig. 5. Convolution structure in MobileNetV3
Fig. 6. Comparison among Relu6, h-swish, and swish activation functions
Fig. 7. Curve of loss function
Fig. 8. Comparison of experiment results by SiameseRPN and improved network
Fig. 9. Comparison of results by different algorithms in four different scenarios
Fig. 10. Experimental results for small targets and target occlusion
Fig. 11. Experimental results for multi-target cross movements
Fig. 12. Experimental results for similar target interference
Fig. 13. Test results of six algorithms on panoramic dataset. (a) Precision; (b) success rate
Classification | Total video | Being blocked | Small object | Similar target | Underexposure | Scale change |
---|
Car | 30 | 3 | 0 | 4 | 23 | 12 | Person | 27 | 7 | 6 | 6 | 1 | 9 | Motor | 12 | 1 | 0 | 1 | 9 | 8 | Bicycle | 6 | 3 | 2 | 0 | 3 | 3 |
|
Table 1. Distribution of tracking difficulties in panoramic data sets
Performance | MDNet | ADNet | RT-MDNet | SiameseRPN | SiameseRPN++ | Ours |
---|
Precision /% | 0.744 | 0.316 | 0.801 | 0.783 | 0.833 | 0.896 | Success rate /% | 0.562 | 0.673 | 0.516 | 0.731 | 0.757 | 0.855 | Speed /(frame·s-1) | 1 | 4 | 8 | 75 | 91 | 33 |
|
Table 2. Performance comparison of all algorithms