Although the dreamviewer tool can visualize and simulate the perception module, it lacks of the ability of visualizing point cloud and ROI area which are helpful for debugging and testing the perception algorithms. For this purpose, we provide an offline visualization tool based OpenGL and PCL libraries to show the obstacle perception result with point cloud. Please see demo videos for this offline perception visualizer on Apollo official website.
We introduce the detailed steps to build and run the offline visualizer in docker as below:
Before running the visualizer, we need to prepare necessary PCD and Pose data which can be extracted from a ROS bag with recorded point cloud data. To facilitate the data extraction, we provide a ROS launch file to dump the PCD frame file and a python script to generate the Pose file for each frame.
1.1 Launch the PCD exporter (see details in Velodyne driver doc)
roslaunch velodyne export_pcd_offline.launch
1.2 Play ROS bag
The default directory of ROS bag is /apollo/data/bag. Assume the file name of ROS bag is example.bag.
cd /apollo/data/bag
rosbag play --clock example.bag
When playing the bag, the PCD files will be dumped to the export directory (defalut: /apollo/data/pcd). The extracted PCD files are named according to their frame numbers which correspond to the order of playing the recorded messages from the point cloud ROS topic (e.g., /apollo/sensor/velodyne64/compensator/PointCloud2). In addition, there are two other files (stamp.txt and pose.txt) generated in the export directory. They will be used to generate the Pose file for each frame.
Notice:
If there are no PCD files exported, please check the bag. It must have PointClouds output, like below:
# rosbag info demo-sensor-demo-apollo-1.5.bag
/apollo/sensor/velodyne64/PointCloud2 1321 msgs : sensor_msgs/PointCloud2
Using 'rosbag info ' to check your bag. In my trial, I use a bag named 'Apollo2.5_demo.bag' (currently named 'demo-sensor-demo-apollo-1.5.bag'), which is 9.59 Gigabytes.
1.3 Generate Pose files
We provide a Python script gen_pose_file.py to generate the Pose files from pose.txt.
cd /apollo/modules/perception/tool
python gen_pose_file.py /apollo/data/pcd
The names of the generated Pose files correspond to their frame numbers as the PCD files. It means that the names of PCD and Pose files for a frame is the same but have different extension names (i.e., .pcd and .pose respectively).
We use Bazel to build the offline perception visualizer.
cd /apollo
bazel build -c opt //modules/perception/tool/offline_visualizer_tool:offline_lidar_visualizer_tool
The option -c opt is used for building the program with optimized performance, which is important for the offline simulation and visualization of the perception module in real time.
If you'd like to run the perception module with GPU, please use the command below:
bazel build -c opt --cxxopt=-DUSE_CAFFE_GPU //modules/perception/tool/offline_visualizer_tool:offline_lidar_visualizer_tool
Before running the visualizer, you may setup the data directories and the algorithm module settings in the configuration file /apollo/modules/perception/tool/offline_visualizer_tool/conf/offline_lidar_perception_test.flag. The detailed parameter settings for each algorithm module can be set according to the corresponding configuration files in /apollo/modules/perception/tool/offline_visualizer_tool/conf/config_manager.config. Then you may run the visualizer with offline perception simulation by the command below:
/apollo/bazel-bin/modules/perception/tool/offline_visualizer_tool/offline_lidar_visualizer_tool
Now you will see a pop-up window showing the perception result with point cloud frame-by-frame. The obstacles are shown with purple rectangle bounding boxes. There are three modes to visualize the point cloud with/without ROI area: (1) showing all the point cloud with grey color; (2) showing the point cloud of ROI area only with green color; (3) showing the point cloud of ROI area with green color and that of other area with grey color. You may press the S key on keyboard to switch the modes in turn.