preface: use limited computing resources and storage resources to achieve low-cost real-time target detection and tracking.
-
-
- darknet
- opencv
- Automatic startup
- remote communication configuration with PC : static IP, ssh
-
-
The current mainstream target detection algorithms are mainly divided into two categories:
- Two stage : Generate Region Proposal (candidate area), and classify the candidate area on this basis.
- One stage :Excluding the candidate region stage of the Two-stage algorithm, the position coordinate value and category probability of the object are directly obtained.Representative algorithms such as YOLO(You only look once)
-
YOLOv3-tiny :
The backbone network has 7 3×3 convolutional layers, 6 pooling layers.
-
DIY leg training dataset.
-
-
Communication via ROS
-
pid tracking:
- pid controller.
-
-
-
-
video as foloows:https://youtu.be/zSw16Put6mo
-
FPS is low, which is about 5-6 fps. The system can only realize real-time tracking at a low speed.
-
-
- limited computing resources
- lacking of virtual memory
-
-
-
In order to provide more resources for the YOLOv3-Tiny detection algorithm to run, it is necessary to increase virtual memory swap and increase 4G virtual memory. At the same time, in order to further reduce the waste of resources, the ubuntu graphical user interface is closed. After finishing these preparations, the test was conducted again, and the number of frames was increased to 20fps. The overall picture was relatively smooth, and it was able to recognize objects in real time.
-
**close the ubuntu graphical user interface **
-