• NVIDIA GPU Driver Version (valid for GPU only):
565.57.01
I am using DeepStream’s line crossing analytics, but sometimes it falsely increments the line crossing count even when no crossing occurs. Below is the configuration I am using:
[line-crossing-stream-2]
enable=1 #Label;direction;lc
#[(676, 275), (672, 244), (578, 236), (778, 236)]
line-crossing-Exit=676;275;672;244;578;236;778;236
class-id=0 #extended when 0- only counts crossing on the configured Line
1- assumes extended Line crossing counts all the crossing
extended=1 #LC modes supported: #loose : counts all crossing without strong adherence to direction #balanced: strict direction adherence compared to mode=loose #strict : stricter direction adherence compared to mode=balanced
mode=loose
I suspect the issue may be related to the mode or configuration settings. How can I prevent these false counts and ensure more accurate line crossing detection? Would adjusting the mode or other parameters help? Any recommendations would be appreciated!
I am not making any modifications to the DeepStream analytics code, and the behavior in question occurs without altering deepstream_app.c or the nvdsanalytics plugin. Most of the time, the line-crossing analytics work accurately, but occasionally, it produces false counts. I’ve provided a reference image below to illustrate the issue. Could you suggest possible reasons for these inconsistencies or any steps to debug and resolve the false positive counts? this is the reference image
I am encountering an issue with false line-crossing counts in DeepStream, even without any modifications to the deepstream_app.c or nvdsanalytics code. While the analytics generally work well, there are occasional false positives. Could you suggest possible solutions or debugging steps to resolve this issue and improve accuracy?
In theory, when the center-bottom of bbox passed the line, the counter will increase. there are other configurations to improve this false positive issue. mode=strict#stricter direction adherence compared to mode=balanced extended=0#If extended is 0, the line crossing only the configured line would be counted
Thank you for the suggestions. I have already set the mode=strict and extended=0 in my configuration under [line-crossing-stream-2] as shown below:
[line-crossing-stream-2]
enable=1
line-crossing-Entry=674;283;678;303;576;307;807;302
class-id=0
extended=0
mode=strict
Despite these settings, I am still encountering occasional false positives in line crossing detection. Today, I also faced a false count. I have attached an image showing an instance of the false count. Could you suggest additional troubleshooting steps or configurations to resolve this issue and improve accuracy?
what the fps and resolution of the original video? could you share the video including the scene of counter increasing? Thanks! wondering the rectangle of bbox. could you share the original video? we want to reproduce this issue. You can use forum private email. please click forum avatar-> personal messages->new message.
In my DeepStream configuration, I set the initial frame resolution in the StreamMux configuration to 1920x1080, but for the analytics section, I set the resolution to 1280x720 using the config-width and config-height parameters.
My question is: Will the line crossing analytics provide correct results with this resolution mismatch (1920x1080 vs. 1280x720)? Or does this difference in resolution between the input stream and the analytics configuration lead to incorrect results or information?
the coordinates in line-crossing-Entry is based on config-width and config-height. low-level will convert the coordinates to a new coordinates in actual frame. in the video shared, I found some bboxes crossed the line.
