Peak Label Object Detection for Radar Position Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems using low-resolution 3D RADAR images face challenges due to inaccuracies in ground truth position information, leading to reduced accuracy in training neural networks and subsequent object detection.
Innovation Solution
The implementation of a peak label object detection (PLOD) system, which uses a fully or semi-supervised segmentor network to output peak shapes indicating the presence of objects, and a peak detector to convert these outputs into detection results, thereby compensating for inaccuracies in object position.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ground truth bounding box information is used to train object detectors, then detection accuracy can be improved, but inaccuracies in ground truth position information deteriorate the training quality and reduce detection performance
Solution Approach 1:
The patent transforms the ground truth annotation from a fixed bounding box into a probabilistic distribution (Gaussian kernel) centered at the object location. This parameter change allows the training data to represent positional uncertainty, where the standard deviation of the Gaussian reflects the expected error margin. The object detector learns to weigh predictions based on this probabilistic ground truth, improving robustness to positioning inaccuracies while maintaining detection accuracy.
2Manufacturing precision
If precise ground truth bounding boxes are manually annotated, then training data quality improves, but labeling effort and time increase significantly
Solution Approach 1:
Instead of requiring annotators to draw precise bounding boxes, the system only requires them to mark the center position of objects. This parameter simplification reduces the annotation task from defining four boundary coordinates to specifying a single positional point, dramatically reducing labeling time while maintaining sufficient training quality through the probabilistic interpretation of the center position.
Solution Approach 2:
The system generates synthetic ground truth labels by copying and transforming simple center position annotations into full probabilistic bounding box distributions using Gaussian kernels. This copying process automatically creates rich training labels from minimal human input, eliminating the need for manual bounding box drawing while preserving annotation quality.
3Ease of manufacture
If traditional bounding box ground truth is used, then training process is straightforward, but detection accuracy suffers due to position inaccuracies in low-resolution images
Solution Approach 1:
The patent changes the ground truth representation from deterministic bounding box coordinates to a probabilistic distribution parameterized by center position and standard deviation. This parameterization naturally handles position uncertainty in low-resolution images by spreading the ground truth probability mass, allowing the detector to learn more robust features without requiring perfect positional precision.
Solution Approach 2:
The Gaussian probability distribution serves as an intermediary between the simple center position annotation and the full bounding box prediction. This intermediary transforms the incomplete annotation into a complete probabilistic label, bridging the gap between minimal human input and comprehensive training requirements while maintaining simplicity.
Data Source
AI summary
A peak label object detection system (PLODS) includes an object size database configured to store information related to object size for a plurality of objects. The PLODS further includes a three-dimensional (3D) sensor database configured to store information related to parameters of a 3D sensor. The PLODS further includes an annotation database configured to store ground truth annotation information for images. The PLODS further includes a peak shape parameter calculator configured to determine a peak label size based on object size from the object size database and the parameters of the 3D sensor. The PLODS further includes a label generator configured to generate a peak labels map based on label size and the ground truth annotation information.