search

Xiangyu-Gao / Raw_ADC_radar_dataset_for_automotive_object_detection

A dataset for the raw ADC data of 2TX-4RX MMWave Radar for automotive object detection.
MIT License
72 stars 16 forks source link

标签数据和Range-angle数据不对齐 #11

Closed luckyyf0517 closed 4 weeks ago

luckyyf0517 commented 6 months ago

标签csv文件中的px、py转换到range-angle坐标系下，在距离稍远的地方出现较大的误差，其中ramap根据雷达参数计算，想请问是自动化标注产生的距离估计误差还是我计算过程出现了问题

Xiangyu-Gao commented 6 months ago

Can you show an example? thanks

luckyyf0517 commented 5 months ago

Thank you for your reply. Here is an example, i read a raw adc file from date sequence 2019_04_09_bms1000, frame 000808.mat, for instance.

The annotation info is 

22,80,-1.11408487,11.31695161,0.6,1.7

We can get px = -1.11 m and py = 11.31 m, which are presented in Cartesian coordinate, so we can calculate the range and angle-of-arrival of the target by $r=\sqrt{x^2+y^2}$ and $a=\arctan{x/y}$ : r=11.37 m, a=-5.62 deg.

Then I used the function labelmap2ra provided in RAMPCNN source code, to map the target into confmap.

Here I draw the ra-view of the radar data, strictly using the mmwave radar parameters provided in https://github.com/Xiangyu-Gao/Raw_ADC_radar_dataset_for_automotive_object_detection/blob/main/config.py, and I draw the target point on the figure as well. (According to the radar config, the resolution of range-bin is about 0.22m. )

image

Obviously there is a disalignemnt between the ramap and the annotation. But in some other frames, when the target is close to the radar, the location of annotation looks correct, for example, frame 000220.mat.

image

I am trying to use the dataset for a semantic segmentation task, but the problem do bother me a lot.

I wonder if I understood the mapping px and py to the RA coordinate correctly, or is there any operation or setting that I missed, or it's simply due to estimation errors caused by semi-automated annotation.

Thank you!

luckyyf0517 commented 5 months ago

I also considered the crop operation mentioned in https://github.com/Xiangyu-Gao/Radar-multiple-perspective-object-detection/blob/aff04b354f07f2689dfd8fcf6fcb68a9404eb73a/utils/mappings.py#L15, but the impact on resolution is small and the results are not improved.

Xiangyu-Gao commented 5 months ago

The crop operation should have little effect on that. I assume this mismatch is an error from the auto-labeling. The further the distance, the larger the error usually.