Fig. 1. Schematic diagram of image acquisition system for runway edge lights
Fig. 2. Structure of the RetinaNet
Fig. 3. Calculation process of standard convolution and depth separable convolution. (a) Standard convolution; (b) depthwise convolution; (c) pointwise convolution
Fig. 4. Structure of the linear inverted residual module. (a) Identity residual block; (b) convolutional residual block
Fig. 5. Structure of the FPN
Fig. 6. Data set image example. (a) Strong natural light image; (b) weak natural light image; (c) image without natural light; (d) image of 1-level light; (e) image of 2-level light; (f) image of 3-level light
Fig. 7. Runway edge light image after data enhancement
Fig. 8. Test results of the test set. (a) Image of 1-level light; (b) image of 2-level light; (c) image of 3-level light; (d) strong natural light image
Fig. 9. Images of runway edge lights with different focal lengths and weather conditions
Fig. 10. Detection results of different models on the same image. (a) Detection results of the model obtained from 3-level light image on 3-level light image; (b) detection results of the model obtained from 1-level light image on 3-level light image
Input | Operator | Output |
---|
DW × DH × M | Conv 1×1,ReLU | DW × DH × tM | DW × DH × tM | DW Conv 3×3,step size is s,ReLU | DW/s × DH /s × tM | DW/s × DH/s × tM | Conv 1×1,ReLU | DW/s × DH/s × N |
|
Table 1. Calculation steps of the inverted residual module
Stage | Operator | Input size |
---|
Stage 1 | Conv 7×7×64,s=2 | 224×224×3 | MaxPool 3×3,s=2 | 112×112×64 | Stage 2 | Conv block | 56×56×64 | identity block×2 | 56×56×256 | Stage 3 | Conv block | 56×56×256 | identity block×3 | 28×28×512 | Stage 4 | Conv block | 28×28×512 | identity block×22 | 14×14×1024 | Stage 5 | Conv block | 14×14×1024 | identity block×2 | 7×7×2048 |
|
Table 2. Structure of the improved feature extraction network
Model | AP(weak)/% | AP(bright)/% | mAP /% | Recall /% | FPS |
---|
SSD | 84.6 | 85.3 | 85.0 | 85.4 | 24.7 | Faster R-CNN | 86.5 | 84.2 | 85.5 | 86.2 | 22.4 | YOLOv4 | 94.3 | 95.5 | 95.6 | 95.8 | 26.5 | RetinaNet | 95.2 | 96.4 | 96.4 | 96.3 | 25.2 | Ours | 96.2 | 97.5 | 97.2 | 96.5 | 25.9 |
|
Table 3. Test results of different models on airport runway edge lights
Data set | AP(weak)/% | AP(bright)/% | mAP /% | Recall /% | FPS |
---|
Strong natural light image | 94.5 | 95.7 | 95.6 | 95.0 | 26.0 | Weak natural light image | 96.2 | 96.7 | 96.6 | 96.0 | 25.6 | Image without natural light | 96.2 | 96.5 | 96.5 | 95.9 | 25.9 | Image of 1-level light | 96.7 | 96.2 | 96.7 | 96.1 | 25.3 | Image of 2-level light | 96.5 | 96.6 | 96.1 | 96.3 | 25.8 | Image of 3-level light | 96.2 | 96.5 | 96.7 | 96.0 | 25.9 |
|
Table 4. Test results of our method on different data sets