Prototypes and results
Prototypes and results
Using the interfaces of the ADORe Perception API, different types of prototypes were developed. Three prototypes were initially examined in more detail: localisation, traffic control and other road users.
In localisation, the position information of the ego-vehicle is implemented. The overriding aim of this prototype is to transmit realistic sensor position information from the simulation to ADORe. This prototype is the basis of the entire system.
During traffic control, traffic control devices such as traffic lights and signs are to be reliably recognised within the simulation. The current objective is limited to the transmission of traffic light data to ADORe.
In the case of road users, all other road users in the vicinity of the ego vehicle must be classified and localised so that the trajectory can be planned. The overarching goal of this prototype is to recognise other road users using the sensor setup created by the user.
Localisation
The first results of the localisation prototype are presented below [as of November 2023]
Current status
- Estimation of the next state with IMU data
- Improvement of position and speed with GNSS data
- Optional: Improvement of orientation with LiDAR sensor data
Implementation
Implementation Summary
Advantages
- Fast results achieved
- Low computing effort
Disadvantages
- Initialisation of the EKF still with ground truth information
- Position determination still too imprecise
Outlook
- Improve localisation
- Carry out initialisation with realistic sensors
- Evaluation through tests
Traffic control
The initial results of the traffic control prototype are presented below [as of November 2023]
Current status
- V2X communication is used
- MAPEN for the digital map of the road
- SPATEM for the status of the traffic light
- OpenDRIVE is used for the position of the traffic lights
- A trained neural network can be used with the camera used
Implementation of
Implementation summary
advantages
- fast results
- Correct data
- Verifiable by drawing the waypoints in CARLA
- Verifiable by drawing the waypoints in CARLA
Disadvantages
- Dependence on CARLA
Outlook
- Design test cases (check reliability)
- Only send ADORe the message for the relevant traffic light
- Traffic light detection via camera
- Confirm V2X messages
Road user
The first results of the road user prototype are presented below [as of November 2023]
Current status
- 3 different solution approaches
- LiDAR-based methods with PointNets and VoxelNets
- Camera-based methods with 3D object recognition and 2D object recognition with depth detection
- Sensor fusion with camera-LiDAR fusion and parallel fusion
Realisation
Summarising
Advantages
- Fast results
- However, problems with integration into ACDC Main (ROS conversion)
- Works with CARLA data
- Neural networks (e.g. PointNets) would have to be retrained on CARLA LiDAR
Neutral
- Calculation instead of prediction
- Vehicle is positioned unfavourably → incorrect calculation
Disadvantages
- State of the art: Neural networks with LiDAR point clouds such as PointNet, VoxelNet, etc.
- Unstable recognition
Outlook
- Design test cases
- Implement another approach (probably PointNet), test and compare with our approach