SIFT-SNN for Traffic-Flow Infrastructure Safety: A Real-Time Context-Aware Anomaly Detection Framework.

Area of Science:

• Computer Vision
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Manual inspection of transportation infrastructure is costly and prone to errors.
• Existing automated systems struggle with motion and lighting variations.
• Neuromorphic systems offer potential for efficient, low-power anomaly detection.

Purpose of the Study:

• To develop an improved neuromorphic vision system for anomaly detection in transportation infrastructure.
• To enhance context-aware and sequence-stable detection using temporal feature aggregation.
• To provide a deployable, interpretable, and energy-efficient alternative to conventional CNN-based inspection tools.

Main Methods:

• Implemented an improved scale-invariant feature transform-spiking neural network (SIFT-SNN) with temporal feature aggregation.
• Encoded SIFT keypoints into latency-based spike trains for classification using a leaky integrate-and-fire (LIF) spiking neural network.
• Evaluated system performance across GPU, CPU, and simulated embedded hardware configurations.

Main Results:

• Achieved 92.3% accuracy and 91.0% macro F1 score with five-fold cross-validation.
• Inference latencies ranged from 9.5 ms to ~48.3 ms per frame across hardware platforms.
• Demonstrated a compact model size (2.9 MB) and low power consumption (5-65 W).

Conclusions:

• The proposed temporally smoothed neuromorphic system offers a robust solution for detecting critical failure modes in infrastructure like barrier pins.
• Temporal smoothing enhances detection recall, particularly for ambiguous cases.
• The system's efficiency and low resource requirements make it suitable for real-world deployment.