A Self-Adaptive Automatic Incident Detection System for Road Surveillance Based on Deep Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Traditional road incident detection systems often lack adaptability and require significant computational resources.
• Real-time monitoring of road conditions is crucial for traffic management and accident prevention.

Purpose of the Study:

• To develop an automatic road incident detection system with low computational complexity for affordable devices.
• To ensure the system is self-adaptive to diverse scenery and road conditions.
• To detect common road incidents including abnormal vehicle speeds, road obstructions, and stopped vehicles.

Main Methods:

• Implementation of lane segmentation and traffic direction identification algorithms.
• Development of techniques to eliminate unnecessary foreground objects for improved focus.
• Utilizing a classifier module with minimal GPU memory (2.3 MBytes) for efficient inference.

Main Results:

• The system processes 80 video frames (400 m coverage) within 12 seconds, detecting vehicles at 120 km/h.
• Achieved accuracy improvements of 2-5% in automatic mode and 2-7% in semi-automatic mode compared to previous methods.
• Demonstrated self-adaptability to various lighting and significant scene changes in real-world scenarios.

Conclusions:

• The proposed system offers a computationally efficient and adaptable solution for automatic road incident detection.
• Its low resource requirements facilitate implementation on low-cost devices, enhancing road safety infrastructure.
• Significant improvements in detection accuracy highlight the system's effectiveness in real-world traffic conditions.