Author Affiliations
1 School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing 100044, China2 Key Laboratory of Vehicle Advanced Manufacturing, Measuring and Control Technology, Beijing Jiaotong University, Beijing 100044, Chinashow less
Fig. 1. Railway scene and track area
Fig. 2. Edge feature map of railway scene
Fig. 3. Distribution of linear character after Hough transformation
Fig. 4. Gaussian convolution kernels rotated by adaptive angles. (a) θ=22°; (b) θ=38°; (c) θ=90°; (d) θ=178°
Fig. 5. Procedures of combining fragmented regions. (a) Strong and weak boundaries; (b) distribution of boundary weight; (c) boundaries after deletion of weak points; (d) fragmented regions; (e) distribution of fragmented region area; (f) local areas after combination; (g)-(o) local areas after segmentation
Fig. 6. Schematic of convolutional neural network structure
Fig. 7. Pre-train convolutional kernels using autoencoder network. (a) Structure of autoencoder networks; (b) pre-trained convolution kernels
Fig. 8. Structural schematic of high-speed railway intrusion detecting system
Fig. 9. Comparison diagrams of results of different algorithms for track area recognition. (a) Railway scenes; (b) manually labeled regions; (c) results of MCG algorithm; (d) results of FCN algorithm; (e) results of proposed algorithm
Kernel size | Kernel quantity | Accuracy /% |
---|
C1 | C2 |
---|
3×3 | 30 | 10 | 72.5 | 100 | 10 | 75.0 | 5×5 | 100 | 10 | 76.0 | 8×8 | 100 | 10 | 76.5 |
|
Table 1. Comparison of experimental results of different CNN network structures
Kernel size | Kernel quantity | Accuracy /% |
---|
C1 | C2 |
---|
3×3 | 30 | 10 | 92.5 | 100 | 10 | 96.0 | 5×5 | 100 | 10 | 98.5 | 8×8 | 100 | 10 | 99.5 |
|
Table 2. Comparison of experimental results of different convolutional neural network structures after optimization
Algorithm | Mean IU /% | Mean PA /% | Mean EP /% | Time /s | Net parameter quantity /106 |
---|
MCG | 72.05 | 79.94 | 10.63 | 7 | — | FCN | 89.83 | 91.26 | 16.20 | 41 | 134 | Proposed algorithm | 81.94 | 95.90 | 18.17 | 2.5 | 0.18 |
|
Table 3. Comparison of experimental results of different algorithms