• https://blog.csdn.net/weixin_45921929/article/details/128656286

1. Using yolov7+tensorrt to achieve target detection

2. Index available GPUs in a multi-GPU environment

3. In a multi-threaded environment, use opencv to put/get pictures into the queue

4. Save the detection pictures and results as jpg and xml to facilitate subsequent iterative training

1.Recommended to use the ubuntu system for operation, and the software configuration is as follows:

• ubuntu20.04
• cuda11.2
• cudnn8.4.0
• tensorrt8.4.3.1
• python3.7
• pytorch1.10.0
• torchvision0.11.0

2.Download this repository

git clone https://github.com/ZhijunLStudio/yolov7_tensorrt_opencv_queue.git

3.Install dependent libraries

pip install -r requirements.txt

1.Modify the detect.py file

# According to your own model and camera information, modify 1, 2, 3
# 1. Put the name of the tensorrt engine under the model folder
trt_name = "best.engine"
# 2. rtsp address, if you are using a USB camera or other onboard camera, you can change it to 0 (without quotation marks)
RtspUrl = "rtsp://admin:[email protected]:554/Streaming/Channels/101"
# 3. Automatically generate xml configuration - tag dictionary, need to follow {"configured folder name": {0: "label 1", 1: "label 2", 2: "label 3"...}} configure
label_dict = {'person': {0: 'person'}}

2.Run the detect.py file

python detect.py

• https://github.com/Monday-Leo/YOLOv7_Tensorrt