reference:
https://www.bilibili.com/video/BV1xL411B7ax
https://www.dgrt.cn/a/2364195.html?action=onClick
https://blog.roboflow.com/how-to-train-yolov8-on-a-custom-dataset/
1. Data set production (generally in coco format):
lableme, or online at https://universe.roboflow.com/
Each data of pictures and labels has a corresponding picture and label, and the two names of picture and label correspond to the same
label. txt file format, the first one is category, the last four are x, y coordinates and w, h
Dataset download address: https://universe.roboflow.com/hero-d6kgf/yolov5-fall-detection/dataset/1
(The roboflow data can basically be used after downloading)
There is only one category.
After downloading, put it in ultralytics-main, download git clone https://github.com/ultralytics/ultralytics.git
Create a dataset folder, put the dataset below, change the name of the dataset, and special symbols will affect
data.yaml
2. Training
Reference: https://www.cnblogs.com/thx2199/p/17165169.html https://mp.weixin.qq.com/s?__biz=MzU1NjEwMTY0Mw==&mid=2247580124&idx=2&sn=d3a2bffed619fb0470de6468a8b7e0e0&chksm= fbc9c0b8ccbe49ae7dffd1fda5aeec8a805df8f227345a7927f9e67c8ac843f734c15d40da08&scene=
27
For training, enter ultralytics-main and open the cmd window:
It is best to use an absolute path for data
yolo task=detect mode=train model=yolov8s.pt data=D:\opencv2\ultralytics-main\dataset\Fall_Detection\data.yaml epochs=15 imgsz=640
CPU training is slow. . .
Breakpoint in training:
Reference: https://blog.csdn.net/qq_40835644/article/details/129283661
复制last.pt的路径，继续训练使用以下命令：
yolo task=detect mode=train model=/home/*****uns/detect/train10/weights/last.pt data=mydata_orange.yaml epochs=300 batch=16 save=True resume=True
The weight of the training process data results will be stored in this address: runs\detect\train2
Model Validation Inference
Training 15epco, the effect feels average
yolo detect predict model=D:\opencv2\ultralytics-main\runs\detect\train2\weights\best.pt source=d5.jpg
