search

koide3 / monocular_person_following

Monocular camera-based person tracking and identification ROS framework for person following robots
192 stars 49 forks source link

`monocular_people_tracking_node.cpp`: All detected persons are not valid. Won't transition (update) from InitialState to InittialTrainingState #21

Open kevin-chau opened 7 months ago

kevin-chau commented 7 months ago

I'm testing monocular person following on the AGX Orin. The webcam and pose estimation seem to be working but it's stuck in the initial state and will not train. I have a full view of my neck and both ankles.

image

So far all I've been able to figure out is that it thinks every person returned by people_tracker->get_people() is invalid. It's failing this check in create_msgs() in monocular_people_tracking_node.cpp:

image

I haven't been able to get a single valid person. I am running on an AGX Orin with CUDA 11.4, tensorflow 1.15.5 (NVIDIA CUDA build), tf-pose 0.1.1. I am able to build everything with catkin_make and no compilation errors.

@koide3 Any suggestions on why all the people detections are invalid? All I can see is that the correction_count() is not greater than validation_correction_count. Thank you so much for your great work on this project!

kevin-chau commented 7 months ago

There are two warnings that stick out to me. One is for the webcam conversion:

image

The other is this ARM64 not supporting NUMA error. I'm not sure what NUMA is.

image

The last thing is that my webcam has a resolution of 768x480.

I'm not sure if any of these factors would effect the tracking algorithm.