search

w111liang222 / lidar-slam-detection

LSD (LiDAR SLAM & Detection) is an open source perception architecture for autonomous vehicle/robotic
Apache License 2.0
498 stars 108 forks source link

How to test performance in a real-world scenario using ROS #18

Open sycoolboy opened 10 months ago

sycoolboy commented 10 months ago

Hello, your work is excellent. I would like to use ROS to test the object detection performance of this code in a real-world scenario. What should I do? I don't have a GPU, and I'm using the Helios lidar and RealSense camera. I'm using the non-CUDA version (auto-ipu) of the code. Docker is already able to receive information from the lidar and camera, but it seems that object detection is not running. Is it because I haven't calibrated it? Below is the terminal information. 2024-01-10 21-36-25 的屏幕截图 2024-01-10 21-36-34 的屏幕截图 2024-01-10 21-36-37 的屏幕截图 2024-01-10 21-36-40 的屏幕截图 2024-01-10 21-36-48 的屏幕截图

w111liang222 commented 10 months ago

hi, the master branch of LSD use the libspconv to realize the sparse 3D conv for object detection, it is built on CUDA, so, the nvidia gpu is necessary.

However, you can test the pointpillar detection model on non-cuda device with the v1.3 tag version. The openvino is used to inference the onnx model on the CPU(very slow, FPS 2~3Hz)

w111liang222 commented 10 months ago

pointpillar model https://github.com/w111liang222/lidar-slam-detection/releases/tag/v1.3.0