Multi Tracking Multi Camera with Re-Identification starts fast and slows down and quits

[Duggy]

Hi,

I am using the :

https://docs.openvinotoolkit.org/latest/omz_demos_multi_camera_multi_target_tracking_demo_python.html

with :

--m_detector FP16\person-detection-retail-0013.xml

and

--m_reid FP32\person-reidentification-retail-0277.xml

and person.py (configs):

sct_config = namespace(
time_window=1, #INSTEAD OF THE 10 Default
continue_time_thresh=2,
track_clear_thresh=3000,
match_threshold=0.25,
merge_thresh=0.15,
n_clusters=4,
max_bbox_velocity=0.2,
detection_occlusion_thresh=0.7,
track_detection_iou_thresh=0.9, 
process_curr_features_number=0,
interpolate_time_thresh=10,
detection_filter_speed=0.6,
rectify_thresh=0.1
)

I am running 3 RTSP cameras. What seems to happen is that everything starts off fine, all the people (busy environment) are detected well. Then after a couple of minutes the frames seems to slow down, the video starts to stutter and a couple of minutes thereafter stops altogether.

Has anyone experienced this or knows a solution?

If I maintain the settings: time_window=10 (instead of 1) then the failure doesn't seem to happen but most/a lot of the people in the view are not identified.

I am running powerful machines (corei9 CPU or corei5 with GPU), processor doesn't climb, memory doesn't alter, still it slows down and fails all the time.

Thanks

[Duggy]

After some investigations, I am led to believe this has to do with the ffmpeg buffer size and openvino (though not proven). The video starts to play at real speed and then slows down to an absolute crawl (person takes a step every 3 seconds type of crawl). Any suggestions?

[IntelSupport]

Hi Duggy,

Thank you for your question. I noticed you have submitted a duplicate ticket in the community. Therefore, I will close this ticket and please wait for further response in this thread: https://community.intel.com/t5/Intel-Distribution-of-OpenVINO/Multi-Tracking-Multi-Camera-with-Re-Identification-crawls/m-p/1310781#M25091

Regards,

Aznie