Target Detection with Road-Surface Context for Non-Whitelist Obstacles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current two-dimensional target detection methods, such as Faster R-CNN and Yolo V3, struggle to accurately detect non-whitelist obstacles on roads, leading to high false detection rates and safety risks in autonomous driving scenarios.
Innovation Solution
A target detection method involving feature extraction on input images to obtain multiple layers of feature maps with varying downsampling rates, merging these maps to generate two-dimensional instance features, and using a road surface segmentation mask to filter out false detections, enhancing accuracy by considering road surface context information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional two-dimensional detectors (Faster R-CNN, Yolo V3) are used for target detection, then the detection process is simple and whitelist obstacle detection is user-friendly, but non-whitelist obstacle detection accuracy is poor and false detection rate is high
Solution Approach 1:
The patent introduces a new dimension by generating a road surface segmentation mask that separates road surface from non-road surface areas. This additional dimensional information (spatial context layer) is overlaid on the traditional 2D detection features, transforming the detection problem from pure 2D coordinate identification to 3D spatial-context identification. The segmentation mask acts as an additional channel or layer that provides contextual information about where obstacles should logically appear, thereby improving non-whitelist obstacle detection without complicating the base detection pipeline.
2Productivity
If conventional detectors are used without road surface context, then detection speed is maintained, but false detection rate exceeds acceptable thresholds (more than one false positive per 100 km)
Solution Approach 1:
The patent performs preliminary action by pre-generating a road surface segmentation mask before the actual target detection process. This mask is created using a trained segmentation model that identifies road surface areas in advance. By preparing this contextual information beforehand, the system can quickly filter detection results against the pre-computed road surface map during runtime, reducing false positives without adding significant computational overhead to the detection speed.
Solution Approach 2:
The road surface segmentation mask serves as an intermediary between the raw detection features and the final detection output. Instead of directly comparing detection boxes with ground truth, the system first checks whether detected targets fall within road surface areas defined by the segmentation mask. This intermediary layer acts as a filter that eliminates false detections (e.g., sky, buildings, trees) while preserving true positive detections on the road surface, thereby improving reliability without sacrificing detection speed.
3Measurement precision
If road surface context information is integrated into detection, then target detection accuracy improves, but system complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the complex detection task into two independent but complementary parts: (1) traditional 2D target detection using existing detectors like Faster R-CNN or Yolo V3, and (2) road surface segmentation using a separate segmentation model. Each component can be trained and optimized independently, and their results are combined through simple spatial overlay. This segmentation approach improves target detection accuracy by adding contextual information while avoiding the need to redesign the entire detection system, thus limiting the increase in system complexity.
Data Source
AI summary
Embodiments of this application provide a target detection method and apparatus, and a storage medium. The method includes performing feature extraction on an input image, to obtain a plurality of layers of feature maps, where the feature map includes road surface feature information, and downsampling rates of the plurality of layers of feature maps are different. The method further includes performing merging on the plurality of layers of feature maps, to obtain a plurality of two-dimensional instance features respectively corresponding to the plurality of layers of feature maps. Furthermore, the method includes performing prediction based on the plurality of two-dimensional instance features, to obtain a road surface obstacle target in the input image. Target prediction is performed based on the feature map including the road surface feature information. Road surface context information is considered, so that accuracy of target detection can be improved.