Road Anomaly Detection and Counting via Vehicle-Speed Frame Skipping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated systems for detecting and counting road anomalies, such as potholes and cracks, face challenges in accuracy, particularly under varying conditions and with overlapping bounding boxes, leading to inefficiencies in road maintenance.
Innovation Solution
A system and method using a machine learning model in an end computing device to detect and classify road anomalies, employing a skipping frame approach based on model fidelity distance and vehicle speed to handle overlapping bounding boxes, and transmitting anomaly notifications to roadside devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If conventional supervised AI systems are used for road anomaly detection, then annotation time and cost are reduced, but accuracy deteriorates in cases of overlapping bounding boxes and subtle anomalies
Solution Approach 1:
The patent segments the anomaly detection process into multiple stages: initial detection using bounding boxes, followed by refinement through pixel-level anomaly score maps. This multi-stage segmentation allows the system to handle overlapping boxes by processing each detected anomaly individually, improving both efficiency and accuracy.
Solution Approach 2:
The patent introduces an intermediary refinement process that takes the initial bounding box detections and enhances them with pixel-level anomaly scoring. This intermediary step acts as a mediator between coarse detection and fine-grained classification, resolving accuracy issues with overlapping boxes while maintaining the efficiency gains of automated detection.
2Reliability
If automated detection systems process all sequential frames, then detection completeness is improved, but computational resources and time are wasted due to repeated detection of the same anomalies
Solution Approach 1:
The patent implements periodic action by processing only key frames at full resolution while using intermediate frames for lighter monitoring. The system detects anomalies in key frames, then uses those detections to guide processing of subsequent frames, creating a periodic pattern of intensive and light processing that maintains completeness while improving efficiency.
Solution Approach 2:
The patent applies preliminary action by detecting and recording anomaly locations in early key frames, then using this preliminary information to guide subsequent frame processing. This allows the system to skip redundant detection of already-identified anomalies while maintaining detection completeness through periodic re-examination of key frames.
3Measurement precision
If high-resolution diverse data is used for training, then model accuracy is improved, but computational requirements increase limiting accessibility
Solution Approach 1:
The patent segments the computational workload by processing images at different resolutions for different purposes: lower resolution for initial anomaly detection and higher resolution only for confirmed anomalies requiring detailed classification. This segmentation reduces overall computational requirements while maintaining model accuracy for critical detections.
Solution Approach 2:
The patent applies partial action by using high-resolution processing only when necessary (for confirmed anomalies requiring detailed classification) rather than applying it uniformly to all images. This selective approach maintains model accuracy for important cases while reducing overall computational requirements to accessible levels.
Data Source
AI summary
A system and a non-transitory computer-readable storage medium for executing a method of detecting road anomalies includes obtaining visual data of a road from an end computing device and inputting the visual data to a machine learning (ML) model that learns to detect and classify at least one road anomaly such as a pothole, longitudinal crack, transverse crack, or alligator crack. The class and a bounding box of at least one road anomaly are output. Multiple post-detection features of the road anomaly, such as an object box area, intersection of neighboring object boxes, union of neighboring object boxes are determined. When the road anomaly is detected across multiple sequential frames, a number of skip frames is determined based on a model fidelity distance (MFD) and a vehicle speed.