Search code examples
blobrosbinary-datapoint-cloud-libraryblobstorage

PointCloud2 Storage Format


I am trying to implement in FPGA a ROS publisher node of PointCloud2 messages. As a first step, i have already implemented a publisher node on the FPGA that is publishing strings. Now, i am trying to do the same but for the PointCloud2 message format.

It is very simple to understand how a string is stored, basically each character is converted to its ASCII value and stored (as it can be seen here). On the other hand, a PointCloud2 is a complex data type that is not so easy to understand.

I have made some progress on understanding how the metadata of a PointCloud2 is stored, however, it is being very difficult to understand the storage of the data part of the PointCloud2 data type. To simplify, I have also tried a PointCloud2 with only one point but i couldn't decode it either. I know that the X, Y and Z are sequentially ordered with 4 bytes each (the datatype is a Float32). Therefore, i can isolate the 4 bytes corresponding to one of the coordinates. I have tried to assign to the X coordinate the values from 0 to 17 (in decimal). This are the values stored when using these values (they are all in decimal):

1 = [0, 0, 128, 63] -> little-endian so, the most significant byte is 63 followed by, 128, 0, 0

2 = [0, 0, 0, 64]

3 = [0, 0, 64, 64]

4 = [0, 0, 128, 64]

5 = [0, 0, 160, 64]

6 = [0, 0, 192, 64]

7 = [0, 0, 224, 64]

8 = [0, 0, 0, 65]

9 = [0, 0, 16, 65]

10 = [0, 0, 32, 65]

11 = [0, 0, 48, 65]

12 = [0, 0, 64, 65]

13 = [0, 0, 80, 65]

14 = [0, 0, 96, 65]

15 = [0, 0, 112, 65]

16 = [0, 0, 128, 65]

17 = [0, 0, 136, 65]

So, my question is, how are the values stored? Supposedly the data part is stored in binary blobs according to here. However, i don't understand what does this mean and how it works. Also, i have not found any concrete example on how to convert a decimal value to this representation.

For a PointCloud2 with X, Y and Z my current understanding is the following (here is the corresponding data):

header:

  • seq (4 bytes)
  • stamp (8 bytes)
  • frame_id (1 byte per character)

height: 4 bytes

width: 4 bytes

fields:

  • number of fields (4 bytes)

field 1

  • dimension (4 bytes)
  • name (1 byte)
  • offset (4 bytes)
  • datatype (1 byte)
  • count (4 bytes)

field 2

  • dimension (4 bytes)
  • name (1 byte)
  • offset (4 bytes)
  • datatype (1 byte)
  • count (4 bytes)

field 3

  • dimension (4 bytes)
  • name (1 byte)
  • offset (4 bytes)
  • datatype (1 byte)
  • count (4 bytes)

is_bigendian: 1 byte

point_step: 4 bytes

row_step: 4 bytes

size: 4 bytes

data: size bytes

is_dense: 1 byte


Solution

  • As I selected the datatype 7 (FLOAT32) the data part is stored as indicated in the standard IEEE 754 (which is a widely used standard to efficiently store floating-point numbers). Thankfully, there is already a Xilinx IP (Floating-Point Operator) to deal with floating-point numbers, including conversions from other types such as INT32 and mathematical operations such as multiplications.