1 Introduction
2 Related works
2.1 Model-based sensor calibration
2.2 Learning-based sensor calibration
2.3 Generative Adversarial Networks (GANs)-based image translation
3 Methodology
3.1 Forward propagation from polar to Cartesian
3.2 Motion distortion of spinning sensors
3.2.1 Motion distortion removal
3.2.2 LiDAR sensor
3.2.3 Imaging radar sensor
3.3 Radar to LiDAR image translation
3.4 Image registration
No | Name | Layer (type) | Output Shape | Parameters |
---|---|---|---|---|
1 | \(\mathsf {c7s1-64}\) | 7\(\times \)7 Convolution-InstanceNorm-ReLU with 64 filters | [-1, 64, 256, 256] | 3200 |
2 | \(\textsf{d128}\) | 3\(\times \)3 Convolution-InstanceNorm-ReLU with 128 filters | [-1, 128, 128, 128] | 73,856 |
3 | \(\textsf{d256}\) | 3\(\times \)3 Convolution-InstanceNorm-ReLU with 256 filters | [-1, 64, 256, 256] | 295,168 |
4 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
5 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
6 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
7 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
8 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
9 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
10 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
11 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
12 | \(\textsf{R256}\) | ResidualBlock with two 3\(\times \)3 Convolutions-InstanceNorm-ReLU | [-1, 64, 256, 256] | 1,180,160 |
13 | \(\textsf{u128}\) | 3\(\times \)3 fractional-strided-Convolution-InstanceNorm-ReLU with 128 filters | [-1, 128, 256, 256] | 295,040 |
14 | \(\textsf{u64}\) | 3\(\times \)3 fractional-strided-Convolution-InstanceNorm-ReLU with 64 filters | [-1, 64, 256, 256] | 73,792 |
15 | \(\mathsf {c7s1-3}\) | 7\(\times \)7 Convolution-InstanceNorm-ReLU with 3 filters | [-1, 1, 256, 256] | 3137 |
No | Name | Layer (type) | Output Shape | Parameters |
---|---|---|---|---|
1 | \(\textsf{C64}\) | 4\(\times \)4 Convolution-InstanceNorm-LeakyReLU with 64 filters | [-1, 64, 128, 128] | 1088 |
2 | \(\textsf{C128}\) | 4\(\times \)4 Convolution-InstanceNorm-LeakyReLU with 128 filters | [-1, 128, 64, 64] | 131,200 |
3 | \(\textsf{C256}\) | 4\(\times \)4 Convolution-InstanceNorm-LeakyReLU with 256 filters | [-1, 128, 64, 64] | 524,544 |
4 | \(\textsf{C512}\) | 4\(\times \)4 Convolution-InstanceNorm-LeakyReLU with 512 filters | [-1, 128, 64, 64] | 2097664 |
4 Experimental results and discussion
4.1 Experimental setup
PSNR | SSIM | |
---|---|---|
Original radar | 14.4038 | 0.0346 |
Median-filtered radar | 15.0587 | 0.0580 |
Gaussian-filtered radar | 14.6781 | 0.0402 |
CNN-filtered radar | 10.3528 | 0.0125 |
Translated radar (10 epoch) | 24.0301 | 0.7319 |
Translated radar (15 epoch) | 24.5550 | 0.7579 |
Translated radar (20 epoch) | 24.3805 | 0.7606 |
Translated radar (25 epoch) | 24.4076 | 0.7781 |
Translated radar (30 epoch) | 24.6909 | 0.7831 |
Translated radar (35 epoch) | 24.6760 | 0.7819 |
Translated radar (40 epoch) | 24.4502 | 0.7829 |
Mean (Extrinsic parameters) | STD (Convergence rate) | |||||
---|---|---|---|---|---|---|
30 epoch | x [cm] | y [cm] | \(\theta \) [rad] | x [cm] | y [cm] | \(\theta \) [rad] |
CycleGAN-PH | 4.4361 | 40.6015 | \(-\)0.0133 | 17.5047 | 24.9289 | 0.1199 |
CycleGAN-MI | \(-\)0.0357 | 56.6839 | \(-\)0.0289 | 12.9081 | 22.4413 | 0.2232 |
CNN-MI | \(-\)25.2390 | 13.5681 | \(-\)0.0076 | 131.9820 | 125.7971 | 0.0775 |
Gaussian-MI | \(-\)59.3272 | 43.3379 | \(-\)0.0086 | 55.3282 | 58.8750 | 0.0095 |
Median-MI | \(-\)63.6261 | 47.5538 | \(-\)0.0093 | 54.1358 | 60.6015 | 0.0093 |
Original-MI | \(-\)55.6915 | 39.8573 | \(-\)0.0082 | 55.7864 | 56.8529 | 0.0097 |