The problem of stereovision-based on-board perception for autonomous mobile systems is complex and imposes challenging constraints on system hardware and software architecture: a wide field of view of the stereo ensemble, high frame rate, dense and accurate data acquisition and processing, robust object detection, tracking and classification.
The solution researched and developed at Technical University of Cluj-Napoca in the Image Processing and Pattern Recognition Center of Computer Science Department is a novel 6D stereovision based perception solution built on two levels architecture and proposing two 3D environment perception methods.
The low level architecture implements the 6D stereovision sensor, providing the primary data for the upper level, consisting of the 3D position and 3D motion information.
Two alternatives are proposed for the high level architecture in charge with the 3D environment description. A structured approach corresponding to a model based detection, tracking and classification of the environment, fit to highly controlled and structured environments. In structured environments, the obstacles are modeled as cuboids, having position, size, orientation and speed.
For the complex or unstructured environments a more general approach called unstructured approach is proposed. The 3D position information provided by the stereovision is used to build a Digital Elevation Map. Elevation map cells are classified into road, obstacles and intermediate generating the Classified Occupancy Maps. Dynamic Occupancy Maps can be generated by associating the 3D motion information to the Classified Occupancy Maps. This set of maps allows the implementation of more effective modeling and inference mechanisms corresponding to the static and dynamic entities from environment.
The main original contributions highlighted in the presentation are the following:
- A new stereo-reconstruction solution using an improved SGM approach (SORT_SGM)
- Original solutions for dense optical estimation
- Ego motion estimation from stereo sequences
- Lane and road geometry perception
- Painted road objects perception
- Obstacles detection, tracking and classification
- Perception and representation of the unstructured environment
- Dynamic Environment Modeling
- Forward collision detection
- High level reasoning on perception and domain knowledge