Occupancy Grid Vehicle Detection With Oriented Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection systems for autonomous vehicles face challenges in achieving high precision and real-time execution with limited computing power, particularly in detecting vehicles in diverse environmental conditions using sensors like LiDARs, radars, and stereoscopic cameras, which have limitations such as degraded performance in rain or limited daytime operation.
Innovation Solution
A method utilizing static occupancy grids and a single-stage neural network architecture for detecting vehicles, which generates a visual representation with oriented bounding boxes, requiring less computational resources and executing in 10 milliseconds, allowing for real-time detection and orientation of vehicles using data from distance sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-stage detection architectures are used to improve detection precision, then measurement precision improves, but computing time and hardware resources increase significantly
Solution Approach 1:
The detection task is segmented into two distinct stages: a coarse detection stage using a lightweight network for rapid candidate identification, and a fine detection stage using a more sophisticated network for precise localization. This segmentation allows the system to achieve high detection precision through the fine stage while maintaining real-time performance through the efficient coarse stage, resolving the contradiction between precision and computing time.
Solution Approach 2:
The coarse detection network performs preliminary detection to identify candidate regions containing objects before the fine detection network processes them. By pre-filtering the search space and only applying computationally intensive fine detection to relevant regions, the system achieves high precision where needed while minimizing overall computing time through this preliminary action.
2Reliability
If complex sensor fusion systems are deployed to improve reliability in diverse environmental conditions, then reliability improves, but device complexity increases
Solution Approach 1:
The detection system is designed to process occupancy grid data from multiple sensor types (LiDAR, radar, cameras) through a unified neural network architecture. The same network structure handles different sensor inputs and various object types (vehicles, pedestrians, cyclists), providing multi-functional capability that improves reliability across diverse environmental conditions while avoiding the complexity of separate specialized systems for each sensor type.
3Productivity
If high-performance computing hardware is used to achieve real-time detection, then productivity improves, but hardware resource consumption increases
Solution Approach 1:
The system employs a dynamic two-stage detection architecture where the computational effort is adaptively distributed between coarse and fine detection stages based on scene complexity. In simple scenes, the lightweight coarse stage provides sufficient detection with minimal resource consumption. In complex scenes requiring higher precision, the system dynamically engages the more resource-intensive fine stage, optimizing the balance between detection speed and hardware resource usage.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a method and associated device for detecting material objects in the space surrounding an autonomous robot. In particular, the invention addresses a method for detecting the presence of vehicles in the environment of an autonomous vehicle, which is based on the analysis of occupancy grids in order to reconstruct, in a top-view visual representation of the autonomous vehicle's environment, bounding boxes oriented around the detected vehicles.