search

strands-project / strands_perception_people

long-term detection, tracking and recognition of people
96 stars 70 forks source link

[bayes_people_tracker] memory leak / lockup #186

Open semeyerz opened 8 years ago

semeyerz commented 8 years ago

We could observe that bayes_people_tracker slows down the whole system at some point by occupying all of the systems memory. Unfortunately I can not provide debugging data(rosbag, logs, ...) yet as the system becomes pretty slow in such situations. There seems to be a correlation between such problems and the number of persons around the robot. This means more people (>8) increase the chance of running into the leak/lockup.

cdondrup commented 8 years ago

First of all, I'm sorry I forgot to answer to your email.

We haven't really observed this and our people tracking is basically running 24/7. Do you know which component exactly is causing the leak? I'll try and investigate this and see if I can reproduce it.

Can you tell me the exact configuration of nodes you are running (e.g. I remember you only ran the UBD?!) and also post the configurations you're using if you changed any of the config yaml files?

semeyerz commented 8 years ago

Thanks for the fast reply. Right now we are only using the laser based leg_detector.

The component causing the leak is bayes_people_tracker (observed with htop).

Right now we use the launch file: https://github.com/CentralLabFacilities/tobi_robot/blob/indigo/tobi_bringup/launch/people_tracker_robot.launch

tlind commented 7 years ago

I know this issue is over a year old, but I can confirm we had a similar issue when we tried out the Bayes tracker in Freiburg. We also observed a correlation with the number of tracked targets. The issue disappeared after switching from NNJPDA to NN data association.

We were trying to debug it for a while and I believe we observed some sort of combinatorial explosion in the NNJPDA code, but couldn't quickly figure out a solution.