pseudolidar-tree-detection

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

pseudolidar-tree-detection

★ 13 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Pseudolidar 3D Obstacle Detection

![Pseudolidar tree detections](https://raw.githubusercontent.com/brian-h-wang/pseudolidar-tree-detection/main/images/detector_example.jpeg)

This repository supports training a 3D object detector to recognize trees, for the purpose of robotic navigation in forests.

We include code for the following:
* Generating labeled 3D object detector training data from ZED stereo camera video. Labeling requires minimal user supervision.
* Training a PointRCNN detector on the labeled data.
* Running a 3D Kalman filter tracker using detections, to filter out detector errors and output stable estimates of tree positions.

## Getting started

### Installation

We've tested our code on Ubuntu 18.04 with Python 3.7.

Dependencies:
- Numpy
- Open3D 0.10 or 0.11 [[Installation instructions](http://www.open3d.org/docs/release/getting_started.html)]

For ZED SVO processing:
- ZED SDK with Python API [[Installation instructions](https://www.stereolabs.com/docs/app-development/python/install/)]

Install PointRCNN according to the instructions in the [PointRCNN repo](PointRCNN/README.md). You'll need to run the following command in the PointRCNN subdirectory:

```git submodule update --init --recursive```

Also install dependencies required by the [AB3DMOT](https://github.com/xinshuoweng/AB3DMOT) repository.

### Data preparation

The `data/...` folders will be created by the code during data processing.

To use our collected ZED SVO data (used in our paper), download the 'training' and 'testing' folders from our [Google Drive folder here](https://drive.google.com/drive/folders/10wYXolnH8tY95gJmOjIu8k3oubIv6AS3?usp=sharing) and add them under the 'zed_svo' directory.

## Data preparation

Currently, this repo supports 3D detector training for SVO format video collected using a Stereolabs ZED camera.

### Recording data

We've tested this repo on video collected using a ZED 2 camera, at 60FPS 720p resolution, using H265 format compression.

### Processing ZED recorded SVO files

ZED SDK and Python API are required for processing SVO files.

Add .svo files to `zed_svo/training` or `zed_svo/testing`. Then, run

``python process_svo_files --svo_dir zed_svo/training --output_dir data/training --n_skip 5``

This processes all .svo files in the `zed_svo/training` directory, creating a for each .svo file a directory containing the following:
- `pointcloud`: Contains sparsified stereo (pseudolidar) pointclouds.
- `rgb`: Contains .jpeg RGB images.
- `map.obj`: Dense fused point cloud global map.
- `poses.txt`: Pose history file.
- `calibration.yaml`: Camera calibration parameters file.

The `n_skip` argument allows skipping every n-th frame in the SVO file, for the pointcloud and rgb outputs.

### Labeling training data

Use the `label_pointcloud_clusters.py` script to manually label point cloud clusters, and use `create_3d_annotations.py` to label point clouds using the clusters.

## Training the object detector

To do - detailed instructions on training.

See the [PointRCNN readme] for instructions on training the 3D detector and performing inference on the test set.

### Convert labeled data to KITTI format

Use `convert_zed_data.py` to convert point clouds, images, and camera calibration info 
 into KITTI dataset format, required by the PointRCNN detector.
### PointRCNN training

Follow instructions in the [PointRCNN readme](PointRCNN/README.md).

## Running the tracker

To run the tracker, use the script `tracking/AB3DMOT/main.py`.

To do - cleaning up tracking code.

## Visualizing results

To visualize tracker results, use

``
python visualize_tracking.py  --results_path output/tracker_results.txt --pcd_dir data/kitti_object/
``

# Work in progress features

- Integrate with other 3D object detectors in addition to PointRCNN
  - To do: Use [mmdetection3d repo](https://github.com/open-mmlab/mmdetection3d) for easier detector training.
- Support for rosbag data from ZED and Intel RealSense

pseudolidar-tree-detection

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README