- A self driving technology is fake if their goal is only to replace manpower and low cost.
- A self driving technology is failure if they do not put safety to the first thing.
- It is time to put self driving technology into engineering and business instead of continuing focus on research and show.
- Autoware is a flexible and powerful open source platform for autonomous vehicles but really hard for beginners because Autoware is nodes oriented not task oriented.
- Autoware and Apollo are all totally based on HD map which generated by Lidar.
2018 GUIYANG MOVE-IT HACKATHON is the first open source hackathon workshop to make real size self-driving car in China, which is organized by a GUIYANG startup company PIX.
There are two cars, one is robot cafe car and another is Honda Civic.
- Can control robot cafe car throttle, brake and steer
- Generate map and waypoints
- Make robot cafe car self-driving follow waypoints
- Detect obstacle with Lidar
- CAN control Civic throttle, brake and steer through comma.ai panda.
- CAN control Civic under low speed.
There are a lot of comma.ai implementation in China with Honda Civic but can not totally control steer under 20km/h.
We completely control Civic even at low speed.
It is pity for limited time we did not totally make Civic self-driving but easy to continue.
- Set up the first open source Chinese Police gesture dataset.
- Train a model through transfer learning with a pre-trained inception model
The Chinese police gesture detection by camera performance is not excellent due to dataset reason and limited time.
More information is shown in Police gestures dataset and detection
Chinese Traffic Gesture Dataset(CTGD)
Please keep mind that this repository is on building now and maybe not complete.
If there are some error when you use ndt_matching with GPU cuda computer:
- Download Autoware develop branch
- Remove the following folder.
Autoware/ros/src/sensing/fusion/packages/autoware_camera_lidar_calibrator
- Compile Autoware develop branch.
Velodyne 32C is a new Lidar while driver and calibration file should be updated.
These files has been saved to Velodyne32C
Copy Velodyne 32C launch file to the following address:
Autoware/ros/src/sensing/drivers/lidar/packages/velodyne/velodyne_pointcloud/launch/32c_points.launch
Copy Velodyne 32C calibration yaml file to the following address:
Autoware/ros/src/sensing/drivers/lidar/packages/velodyne/velodyne_pointcloud/params/VLP-32C.yaml
Copy Velodyne 32C driver cc file to the following address:
Autoware/ros/src/sensing/drivers/lidar/packages/velodyne/velodyne_driver/src/driver/driver.cc
Velodyne 32C factor distance to resolution is different from others Velodyne products, this factor should be modified as following:
- driver file
Autoware/ros/src/sensing/drivers/lidar/packages/velodyne/velodyne_pointcloud/src/lib/rawdata.cc
- Change
float distance = tmp.uint * DISTANCE_RESOLUTION;tofloat distance = tmp.uint * 0.004;
- Install ros-velodyne driver:
sudo apt-get install ros-VERSION-velodyne
- Connect velodyne 32C and disconnect wifi
- Velodyne 32C IP setting
- Robot Cafe car: 192.168.1.201
- Civic: 192.168.0.201
- Computer IP set with in Lidar net, e.g. (robot cafe car 192.168.1.100; Civic 192.168.0.100)
- View Lidar data
roslaunch velodyne_pointcloud 32c_points.launch
rosrun rviz rviz -f velodyne
ROS Velodyne driver install and connect guide
The computer ip should be set in Lidar net but last not 201 (which is for Lidar).
The Lidars ip can be set same ip.
Make sure you have connected with Lidar successfully and have enough free disk space.
roslaunch velodyne_pointcloud 32c_points.launch
rosrun rviz rviz -f velodyne
rosbag record -a
- Load simulation rosbag file, play and pause.
- Change rosbag topic name to /points_raw
- The demo.rosbag can be used here for following tutorial.
The above two step can be done with the following commands:
rosbag play -r 0.7 bag_name.bag /velodyne_points:=/points_raw
You can use space button to stop and play rosbag in ternimal.
- Downsample rosbag files with voxel_grid_filter.
When you click ROSBAG Record stop button, the new downsample rosbag will be saved.
- Change downsample rosbag topic name to /points_raw
rosbag play -r 0.7 bag_name.bag /filtered_points:=/points_raw
- Active ndt_localizer | ndt_mapping
- waypoint_maker | waypoint_saver
- Run whole simulation rosbag
- Output pcb
- Save waypoints
- The Velodyne default topic name is velodyne_points.
- The downsample rosbag default topic name is filtered_points/
- Please confirm voxel_grid_filter and ROSBAG Record topic name agree with rosbag playing.
- You can check topic data available or not using Autoware Topics.
- Make sure modify app parameters first then active related nodes function.
Here is simulation process and rviz visualization with generated pcb and waypoints file.
- Setup, sensing
- Load Map load pcb, waypoints file
- Computing setting
- Rviz to simulate
- If error please check every settings and redo it again.
- Make sure active vel_pose_connect in Simulation Mode.
- Make sure waypoint_follower is waypoint.
- Make sure deactive vel_pose_connect Simulate Mode.
- Make sure deactive waypoint_follower | wf_simulator.
- Make sure velocity_set | Points_topics is points_no_ground.
- Autoware Lidar obstacle detection function will false detect upslope as obstacle and don't move.
- The dataset should be big and diversity enough to prevent deep learning model over fitting.
- LSTM model has been used instead of CNN model to consider time serial.
- Police gesture detection have been localized for different countries.
- Through we have control robot cafe car through CAN, the driving performance is really bad, especially on upslope.
- There are still a lot of improvement potential for vehicle OEM and Tier1 e.g. Bosch to do in the hardware and CAN control.
The specific code is distributed under MIT License.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent