Background
Long-range detection of objects and dangers in the road is an ongoing challenge for the field of vehicle autonomy. For an automated vehicle, the distance to detect another car or an obstacle in the road directly correlates to the maximum speeds the vehicle can safely drive and react. Radar is one of the most common methods for long distance detection, but the data is noisy (from reflections, multi-path, and limited field of view) and radars often provide less resolution than some alternative sensing modalities. Alternative sensors, such as Lidar and cameras, independently cannot replicate the velocity measurement and range capability of radar. We believe augmenting radar with the spatial information from Lidars and cameras will increase radar reliability to the level necessary for autonomous vehicle operation. To solve the identified problems, we propose to evaluate the effectiveness of combining technologies to reduce the ambiguities associated with any single long-range object detection technology.
Approach
Our approach was to:
- Create Lidar algorithms for long-range object detection, using 2D road info and radar target initialization
- Register detected targets to pre-mapped road lanes using 2D road info and precise localization
- Output a final augmented radar detection that can be used for making decisions by higher level autonomy
Accomplishments
Our accomplishments thus far include:
- Collected ground truth data using a pair of vehicles for high precision and for arbitrary vehicles observed on collection runs
- Designed software to match/track detections over time with an extended kalman filter
- Created software to associate detections with pre-mapped road lanes and predict object motion along roads, even when losing sight of objects