VL53L8 application help, looking specifically at delta between data sets

[RSarv]

Hello community,

I'm new to using this sensor, but based on some initial testing I'm hoping it's a good fit for my application. I was hoping to get some general opinions/guidance on the approach I'm planning, so any tips/feedback are greatly appreciated!

My application:

Collect some reference data for a small (~1.5x1.5m) room, with the sensor mounted fixed in the middle of the ceiling, about 2.75m from the floor, looking straight down to the room. There are some stationary objects in the room that will be part of this reference 'scene' that is being captured. It is a fully enclosed room (expected to have no sunlight when the door is closed), and we can turn the lights off, so there should be little to no ambient light.

I'd like to be able to periodically rescan the room and look for any differences to the reference 'scene' -- essentially looking for unexpected objects. It is not a very time sensitive application, so any filtering etc that might be helpful in making this determination is fine (within reason).

I did some initial, very basic testing using the STSW-IMG041 application to collect logs, and then do some post-processing with them. I applied the same processing to both the reference data (empty room) and data with a backpack in the room. I was using 8x8 zone configuration, 4Hz, and I captured about 10s of data for each log. I tested a couple different basic filtering approaches (where this filtering is done on a zone-by-zone basis across the whole data log), and was able to clean up the data fairly well. Then, I take the filtered result for each data set (which ends up as a single value for each of the 64 zones), take the difference between the two data sets, and generate a depth map of that result. So in a perfect world, that depth map would be 0 for each zone where no extraneous object is present. Of course in reality this is not the case, due to the noise in the data.

I placed the backpack on a section of the floor where there are none of the expected permanent objects in those zones (in other words, those zones should be almost completely comprised of backpack), and I'm able to cleanly see the backpack in this test. The delta depth map result is something like 45mm, whereas the actual depth of the backpack was probably something more like 150mm, so the accuracy is not there, but I actually don't care about that for this application -- I just care that the result is high enough above the noise floor to confidently recognize it as an extraneous object.

So speaking of that 'noise floor', for zones where only the floor of the room is present (and therefore those zone should only be seeing one depth result), the delta depth map is actually surprisingly clean; <2mm across the board. But the zones with multiple depth readings in them give quite a bit more 'noise': up to about 13mm.

So I've got a couple main questions here. One is regarding what the fundamental limitation of the smallest (depth-wise) detectable object might be (I know this will likely depend on the reflectivity of that object). I did a test with a ~50mm tall object (black and non-reflective, to be fair) and it was just lost in the noise floor. I'm wondering if the max ranging capabilities and range accuracy info (tables 19 and 20) in the sensor datasheet can be used for this assessment. I guess this is fundamentally a question about what 'accuracy' means in this context. Because I don't care about accuracy strictly speaking, in that each measurement could be 50mm off and I wouldn't care; but if the repeatability is really good then I still get an accurate delta. But if the accuracy spec in Table 20 is saying you can get that level of variation between multiple readings with otherwise fixed measurement conditions, then that would mean for example a gray target in the dark with 8x8 zones is +/-5% and that corresponds to +/-137mm. Should I be using that as an estimate for the smallest gray object I could detect, or is it more nuanced than that?

I'm also wondering if digging into the raw histogram data is merited in this case, specifically to allow for detection of smaller objects?

My other question is regarding those zones where there are multiple depth results in the reference data set. To detect extraneous objects in those zones, I imagine looking at the raw histogram data may allow for more nuanced extraction of a good delta value?