Radar obstacle detecting system and method thereof

Abstract

This disclosure provides a radar obstacle detection method, including a radar detection step, a descrambling step, and an obstacle information determination step. The radar detection step uses a radar installed in a vehicle to detect obstacles. The descrambling step involves the vehicle's processor receiving the echo signal returned by the radar. The processor descrambles the echo signal using a modified AlexNet model. The modified AlexNet model includes an input layer, five original convolutional layers, two residual convolutional layers, three batch normalization layers, nine excitation layers, a stacking layer, and three fully connected layers. In the obstacle information determination step, the processor determines the distance and angle of the obstacle based on the descrambled echo signal to improve detection accuracy.

Claims

1. A radar obstacle detection method, comprising: a radar detection step, wherein a radar installed in a vehicle detects an obstacle; and a descrambling step, wherein a processor of the vehicle receives an echo signal transmitted from the radar, the processor descrambling the echo signal using a modified AlexNet model, the modified AlexNet model comprising: an input layer for receiving the echo signal; a first to a fifth original convolutional layer, wherein the first original convolutional layer is directly connected to the input layer; and a first to a second residual convolutional layer connected between a second and a third original convolutional layer among the first to the fifth original convolutional layers. The system comprises three batch normalization layers: the first batch normalization layer directly follows the first original convolutional layer; a second batch normalization layer directly follows the second original convolutional layer; and a third batch normalization layer directly follows the first residual convolutional layer. The system also comprises nine activation layers: the first activation layer connects the first batch normalization layer and the second original convolutional layer; a second activation layer connects the second batch normalization layer and the first residual convolutional layer; a third activation layer connects the third batch normalization layer and the second residual convolutional layer; and a fourth activation layer connects the first batch normalization layer and the second residual convolutional layer. The activation layer is connected before the third original convolutional layer. The fifth activation layer among the first to the ninth activation layers is connected between the third original convolutional layer and the fourth original convolutional layer among the first to the fifth original convolutional layers. The sixth activation layer among the first to the ninth activation layers is connected between the fourth original convolutional layer and the fifth original convolutional layer. The seventh activation layer among the first to the ninth activation layers is directly connected to the fifth original convolutional layer. A stacking layer that receives the outputs of the second original convolutional layer and the second residual convolutional layer and is connected before the fourth excitation layer; and the first to third fully connected layers, wherein the first fully connected layer is connected between the seventh excitation layer and an eighth excitation layer among the first to ninth excitation layers, the second fully connected layer among the first to third fully connected layers is connected between the eighth and ninth excitation layers, and the third fully connected layer is directly connected to the ninth excitation layer; and an obstacle information determination step, wherein the processor determines a distance and an angle of the obstacle based on the deduplicated echo signal.

2. The radar obstacle detection method as described in claim 1, wherein each of the first to fifth original convolutional layers contains a complex number of one-dimensional convolutional kernels.

3. The radar obstacle detection method as described in claim 2, wherein the number of one-dimensional convolutional kernels in the first original convolutional layer and the second original convolutional layer is the same.

4. The radar obstacle detection method as described in claim 1, wherein the radar includes a frequency-modulated continuous wave chip.

5. The radar obstacle detection method as described in claim 1, wherein the improved AlexNet model further includes a regression output layer following the third fully connected layer.

6. A radar obstacle detection system, comprising: a radar for detecting an obstacle and transmitting an echo signal; and a vehicle for mounting the radar thereon and including a processor, the processor receiving the echo signal transmitted by the radar, the processor including a modified AlexNet model to scramble the echo signal, and using the scrambled echo signal to determine a distance and an angle of the obstacle, wherein the modified AlexNet model includes: an input layer for receiving the echo signal; first to fifth original convolutional layers, wherein the first original convolutional layer is directly connected to the input layer; and first to second residual convolutional layers connected between a second and a third original convolutional layer among the first to fifth original convolutional layers. The system comprises three batch normalization layers: the first batch normalization layer directly follows the first original convolutional layer; a second batch normalization layer directly follows the second original convolutional layer; and a third batch normalization layer directly follows the first residual convolutional layer. The system also comprises nine activation layers: the first activation layer connects the first batch normalization layer and the second original convolutional layer; a second activation layer connects the second batch normalization layer and the first residual convolutional layer; a third activation layer connects the third batch normalization layer and the second residual convolutional layer; and a fourth activation layer connects the first batch normalization layer and the second residual convolutional layer. The activation layer is connected before the third original convolutional layer. The fifth activation layer among the first to the ninth activation layers is connected between the third original convolutional layer and the fourth original convolutional layer among the first to the fifth original convolutional layers. The sixth activation layer among the first to the ninth activation layers is connected between the fourth original convolutional layer and the fifth original convolutional layer. The seventh activation layer among the first to the ninth activation layers is directly connected to the fifth original convolutional layer. A stacked layer that receives the outputs of the second original convolutional layer and the second residual convolutional layer and is connected before the fourth activation layer; and a first to a third fully connected layer, wherein the first fully connected layer is connected between the seventh activation layer and an eighth activation layer among the first to the ninth activation layers, a second fully connected layer among the first to the third fully connected layers is connected between the eighth activation layer and the ninth activation layer, and the third fully connected layer is directly connected to the ninth activation layer.

7. The radar obstacle detection system as claimed in claim 6, wherein each of the first to fifth original convolutional layers contains a complex number of one-dimensional convolutional kernels.

8. The radar obstacle detection system as described in claim 7, wherein the number of one-dimensional convolutional kernels in the first original convolutional layer and the second original convolutional layer is the same.

9. The radar obstacle detection system as described in claim 6, wherein the radar includes a frequency-modulated continuous wave chip.

10. The radar obstacle detection system as described in claim 6, wherein the modified AlexNet model further includes a regression output layer following the third fully connected layer.

Images