tensorrt_bevdet is a dynamic 3D bev object detection package based on 6 surround view cameras.
BEVDet is a BEV perception algorithm based on panoramic cameras. It unifies multi-view images into the perspective of BEV for 3D object detection task. In this implementation, BEVDet network to inference with TensorRT.
| Name | Type | Description |
|---|---|---|
~/input/pointcloud |
sensor_msgs::msg::PointCloud2 |
input pointcloud (only used for time alignment and display) |
~/input/topic_img_fl |
sensor_msgs::msg::Image |
input front_left camera image |
~/input/topic_img_f |
sensor_msgs::msg::Image |
input front camera image |
~/input/topic_img_fr |
sensor_msgs::msg::Image |
input front_right camera image |
~/input/topic_img_bl |
sensor_msgs::msg::Image |
input back_left camera image |
~/input/topic_img_b |
sensor_msgs::msg::Image |
input back camera image |
~/input/topic_img_br |
sensor_msgs::msg::Image |
input back_right camera image |
| Name | Type | Description |
|---|---|---|
~/output/boxes |
autoware_perception_msgs::msg::DetectedObjects |
detected objects |
~/output/pointcloud |
sensor_msgs::msg::PointCloud2 |
output pointcloud (only used for display) |
The model is trained on open-source dataset NuScenes and has poor generalization on its own dataset, If you want to use this model to infer your data, you need to retrain it.
You can download the onnx format of trained models by clicking on the links below.
- BEVDet: bevdet_one_lt_d.onnx
The model was trained in NuScenes database for 20 epochs.
If you want to train model using the TIER IV's internal database(~2600 key frames), please refer to the following repositories:BEVDet adapted to TIER IV dataset
[1] https://github.com/HuangJunJie2017/BEVDet/tree/dev2.1