Multi-Domain Convolutional Neural Network Tracking Algorithm Based on Reconstructed Feature Combination

Dawei Yang; Xinfei Gong; Lin Mao; Rubo Zhang

doi:10.3788/LOP56.191501

Laser & Optoelectronics Progress, Volume. 56, Issue 19, 191501(2019)

Multi-Domain Convolutional Neural Network Tracking Algorithm Based on Reconstructed Feature Combination

Dawei Yang^1,2, Xinfei Gong^1、*, Lin Mao^1,2, and Rubo Zhang^1,2

Author Affiliations

¹College of Mechanical and Electronic Engineering, Dalian Minzu University, Dalian, Liaoning 116600, China

²Key Laboratory of Intelligent Perception and Advanced Control State Ethnic Affairs Commission, Dalian Minzu University, Dalian, Liaoning 116600, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. MDNet network structure diagram

Download full size

Fig. 2. Feature visualization results (taking some features as examples). (a) Input image; (b) features of Conv3 layer

Download full size

Fig. 3. Deconvolution implementation process

Download full size

Fig. 4. Visualization results of reconstructed features (taking some features as examples). (a) Input image; (b) features of Conv3 layer; (c) reconstructed features

Download full size

Fig. 5. Feature combination

Download full size

Fig. 6. RCNet network structure (?? indicates feature combination)

Download full size

Fig. 7. Feature combination analysis. (a) Input image; (b) features of Conv3 layer; (c) features of Conv5 layer; (d) combination features

Download full size

Fig. 8. OTB50 experimental results. (a) Tracking precision score; (b) tracking success score

Download full size

Fig. 9. Maps of tracking success rate attributes. (a) Low resolution; (b) background clutter; (c) deformation; (d) occlusion; (e) scale variation; (f) out-of-plane rotation; (g) motion blur; (h) illumination varition

Download full size

Fig. 10. VOT2015 testing results

Download full size

Table 1. Average tracking speed of various deep learning traking algorithms based on OTB50
View table
Table 1. Average tracking speed of various deep learning traking algorithms based on OTB50
Algorithm CNN-SVM MDNet TCNN RCNet
Tracking speed /(frame·s^-1) 1.45 1.21 2.14 1.06

Tools

Get Citation

Copy Citation Text

Dawei Yang, Xinfei Gong, Lin Mao, Rubo Zhang. Multi-Domain Convolutional Neural Network Tracking Algorithm Based on Reconstructed Feature Combination[J]. Laser & Optoelectronics Progress, 2019, 56(19): 191501

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites