Abnormal Vehicle Detection via Background Modeling in Road Video
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based methods for detecting abnormally stopped vehicles face challenges due to the small number of samples and insufficient accuracy in sample labeling, leading to poor performance in identifying abnormal stopping behavior in unknown scenes.
Innovation Solution
A detection method that involves obtaining surveillance videos, performing background modeling to extract background images, conducting vehicle detection on these images, and using Intersection over Union (IoU) to identify detection boxes of abnormal vehicles with abnormal stop events, along with differential mask extraction and image sequence analysis to refine the detection process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If deep learning-based methods are used to detect abnormally stopped vehicles, then detection capability is provided, but accuracy deteriorates due to limited samples and insufficient labeling
Solution Approach 1:
Instead of training the model to detect abnormal stopping behavior directly (which requires scarce abnormal samples), the patent inverts the approach by training on normal driving behavior samples and detecting deviations from this normal pattern. This allows the system to achieve high accuracy in abnormal detection by leveraging the abundance of normal driving data rather than relying on scarce abnormal samples.
Solution Approach 2:
The patent changes the training parameter from direct abnormal behavior detection to normal behavior pattern learning. By modifying what the model learns (from abnormal patterns to normal patterns), the system can achieve better detection accuracy despite the scarcity of abnormal samples, as the model can leverage extensive normal driving data for training.
2Quantity of substance
If normal driving vehicle samples are used to train detection models, then training data availability is improved, but detection accuracy for abnormal events deteriorates
Solution Approach 1:
The patent inverts the traditional training approach by using normal driving samples to train the model and then detecting abnormalities as deviations from these normal patterns. This allows the system to utilize the abundant normal driving data for training while maintaining high accuracy in detecting rare abnormal events, as the model learns what normal behavior looks like and can identify deviations from this baseline.
Solution Approach 2:
The patent introduces an intermediary approach where normal driving behavior serves as a reference baseline. The detection mechanism uses this normal behavior reference to identify abnormal events, effectively mediating between the abundant normal training data and the scarce abnormal detection requirement. This intermediary reference allows the system to leverage training data availability while maintaining detection accuracy.
3Device complexity
If detection models are trained on limited abnormal samples, then model training complexity is reduced, but detection reliability in unknown scenes deteriorates
Solution Approach 1:
The patent changes the training parameter from learning specific abnormal patterns to learning normal behavior patterns. This parameter change allows the model to be trained on abundant normal driving data, which improves reliability in unknown scenes by teaching the model what normal behavior looks like across diverse conditions, rather than relying on limited and potentially biased abnormal samples.
Solution Approach 2:
The patent performs preliminary training on normal driving behavior to establish a comprehensive baseline of what normal operation looks like under various conditions. This preliminary action creates a robust foundation that enables the model to reliably detect abnormalities in unknown scenes, as the model has already learned the full range of normal variations before encountering abnormal events.
Data Source
AI summary
This disclosure provides a detection method and an apparatus of abnormal vehicle, device, and storage medium, wherein the detection method includes: obtaining a surveillance video of a target road; performing background modeling based on the surveillance video to obtain background images of at least partial video frames in the surveillance video; performing vehicle detection processing on the background images; and determining detection boxes located on the target road in the background images as detection boxes of abnormal vehicles with abnormal stop events.