Ask Your Question
0

Single Object Tracking in RealTime without cuda

asked 2020-11-24 01:31:19 -0600

Hi, searching online i can't understand what to choose for tracking single object using low hardware like raspberry with edge tpu. I read that SiamFCpp is the latest tested, but i see that it require cuda gpu. Then i found many code like SiamDW, OceanPlus,MegviiDetection/video_analyst but all of that require cuda gpu and i'm so confused on what to do or what to learn. I cannot use nvidia gpu so i think that i need to search something that using edge tpu usb or intel movidius but cannot get something working. I try to train custom object detection using google colab and i get a .pth weights file that i can use to detect custom object but using cpu and webcam the fps is very slow, like 1fps or so.

What i'm doing wrong? Could someone suggest me or route me what to read or so to get single object tracking in realtime.

Thanks

Thanks

edit retag flag offensive close merge delete

Comments

tracking and detection are different tasks. tracking of arbitrary objects can be done reasonably well without any pre-trained model. detection can be used to find an object and put a tracker on it. if your object class is somewhat common, you can just pick one of the commonly used object detection networks (trained weights are available). if you need a specific object class, you can take an existing net and weights, and retrain its classification layer, which is a lot less work.

crackwitz gravatar imagecrackwitz ( 2020-11-24 09:19:11 -0600 )edit

1 answer

Sort by ยป oldest newest most voted
2

answered 2020-11-27 03:13:58 -0600

You are on the OpenCV forum, so I guess you want to use OpenCV functionality for visual object tracking. In OpenCV, you can use Python 3 samples of DaSiamRPN and SiamRPN++ tracker: https://github.com/opencv/opencv/tree... Or you can use trackers from the opencv-contrib "tracking" module: https://github.com/opencv/opencv_contrib You can use them without GPU. But I did not try to run them with Raspberry or something like that.

edit flag offensive delete link more

Question Tools

1 follower

Stats

Asked: 2020-11-24 01:31:19 -0600

Seen: 3,100 times

Last updated: Nov 27 '20