Road Traffic Incident Detection with Small-Sample Twin Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional traffic incident detection technologies face challenges due to insufficient incident samples, leading to high missing report rates and model insensitivity, and limited sample numbers causing over-fitting and insufficient generalization.
Innovation Solution
A road traffic incident detection method utilizing small sample learning, involving data collection, feature extraction, sample pairing, and a twin network architecture to construct a traffic incident detection model, with non-incident samples down-sampled through time series clustering and paired with incident samples to enhance model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are used to detect traffic incidents, then detection accuracy is improved, but the model becomes insensitive to few samples and has high missing report rate due to severe class imbalance
Solution Approach 1:
The patent segments the training process into two stages: first training on paired samples (incident + corresponding non-incident), then on unpaird samples. This segmentation allows the model to learn from balanced pairs first, preventing insensitivity to few samples, while the second stage generalizes to all data, reducing missing report rates.
Solution Approach 2:
The patent performs preliminary pairing of incident and non-incident samples before training. By pre-establishing corresponding sample pairs, the model receives balanced training data initially, which prevents it from being insensitive to few samples, while still maintaining the ability to detect all incident types.
2Adaptability or versatility
If more non-incident samples are added to balance the dataset, then model generalization is improved, but the total number of samples increases and training becomes more complex
Solution Approach 1:
The patent performs preliminary pairing of incident and non-incident samples before training. By pre-establishing corresponding sample pairs, the model receives balanced training data initially, which prevents it from being insensitive to few samples, while still maintaining the ability to detect all incident types.
Solution Approach 2:
The patent segments the training process into two stages: first training on paired samples (incident + corresponding non-incident), then on unpaird samples. This segmentation allows the model to learn from balanced pairs first, preventing insensitivity to few samples, while the second stage generalizes to all data, reducing missing report rates.
3Ease of manufacture
If manual threshold setting is used for incident detection, then implementation is simple, but the threshold setting is subjective and has low incident identification rate
Solution Approach 1:
The patent replaces manual threshold setting with an automated deep learning model that learns from data. The model automatically identifies incident patterns through training on paired and unpaird samples, eliminating subjective threshold selection while achieving high incident identification rates through data-driven decision making.
Data Source
AI summary
Provided are a road traffic incident detection method, system and device based on small sample learning. According to the invention, traffic flow data and incident data are utilized to construct non-incident samples by a case contrast study method, traffic flow feature indexes are extracted to construct incident and non-incident sample feature sets, then sample division and pairing are performed on the feature sets to obtain training and test sample pair sets, a traffic incident detection model is constructed by adopting a twin network architecture in small sample learning, and the model is trained and tested by a sample pair mode. The invention is conductive to accurately detecting accidental traffic incidents, and providing supports for avoiding secondary accidents and improving the traffic safety level.