Autonomous Object Detection With Adaptive Polygon Bounding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection systems for autonomous and semi-autonomous systems suffer from limited accuracy and computational inefficiencies, particularly in real-time applications, leading to suboptimal navigation and movement in complex environments.
Innovation Solution
An object detector using a deep neural network (DNN) predicts a variable number of vertices for bounding polygons, employing a deformable polar coordinate system to enhance detection precision and reduce inference latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If rectangular bounding boxes are used for object detection, then the detection process is simple and fast, but the shape and location accuracy of detected objects deteriorates
Solution Approach 1:
The patent segments the object detection task into two stages: first detecting objects with bounding boxes, then refining the detection by predicting polygon vertices for each detected object. This segmentation allows the system to maintain the speed advantage of bounding box detection while achieving the precision of polygon-based methods.
Solution Approach 2:
The system performs preliminary object detection using bounding boxes before refining the detection with polygon vertex prediction. This preliminary action enables the system to quickly identify potential objects of interest, which are then processed in greater detail, improving overall efficiency and accuracy.
2Shape
If ground truth polygons with evenly-spaced angular rays are used, then object shape representation improves, but detection accuracy deteriorates due to under/over prediction
Solution Approach 1:
The patent transforms the static, evenly-spaced angular ray system into a dynamic one where the number and positioning of rays are adaptive. The system learns optimal ray configurations during training, allowing rays to be denser in regions requiring higher precision and sparser in simpler regions, thus improving detection accuracy while maintaining computational efficiency.
Solution Approach 2:
The system changes the parameters of the ray system from fixed evenly-spaced angles to learned adaptive angles and densities. By allowing the number of rays and their angular positions to vary based on object characteristics, the system achieves better shape representation and detection accuracy without the limitations of uniform sampling.
3Measurement precision
If pixel-level labeling is used for instance segmentation, then detection accuracy improves, but computational cost and processing time deteriorate
Solution Approach 1:
The patent segments the computational task by avoiding full pixel-level segmentation. Instead, it detects objects with bounding boxes and then predicts only the essential polygon vertices that define object shapes. This segmentation of the labeling task dramatically reduces computational complexity while maintaining sufficient accuracy for autonomous navigation.
Solution Approach 2:
The system extracts only the critical geometric information needed for object detection and navigation—specifically, the vertices of bounding polygons—rather than processing all pixel-level data. This extraction of essential features achieves accurate object representation without the prohibitive computational cost of complete instance segmentation.
4Loss of time
If a fixed number of vertices is used for polygon prediction, then inference latency is reduced, but detection flexibility and accuracy deteriorate
Solution Approach 1:
The patent implements a dynamic vertex prediction system where the number of vertices predicted for each object is not fixed but adapts based on object complexity. The system can predict a variable number of vertices per object, allowing simple objects to use fewer vertices (faster inference) while complex objects receive more vertices (higher accuracy), thus resolving the trade-off between speed and flexibility.
Data Source
AI summary
In various examples, one or more object detectors may regress bounding polygons for detected objects in systems (e.g., autonomous or semi-autonomous driving systems and applications) that provide object awareness, object identification, object avoidance, and/or object localization. The object detector may determine regression data representing a regressed polygon associated with a given shape of a detected object represented by classification data determined from a scene. The object detector may determine regression data for different regressed angles between different pairs of successive vertices of the regressed polygon and regressed lengths of vectors from a regressed geometric center of the regressed polygon to vertices of the regressed polygon. The object detector may generate, based at least in part on the regression data, a bounding shape for a detected object in the scene. In some embodiments, the object detector may be trained by deforming a regressed polygon to match a ground truth polygon.