Multi-object detection at night for traffic investigations based on improved SSD framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Safety Engineering

Background:

• Nighttime traffic object detection is challenging due to poor lighting, impacting traffic investigations.
• Existing vision-based methods struggle with medium and small stationary objects at night.
• Accurate detection of traffic objects is crucial for incident analysis and road safety.

Purpose of the Study:

• To develop an effective nighttime multi-object detection framework for traffic investigations.
• To enhance the detection of medium and small stationary objects under low-light conditions.
• To improve the overall performance of traffic object detection at night, particularly at intersections.

Main Methods:

• A nighttime multi-object detection framework was developed using the Single Shot MultiBox Detector (SSD) architecture.
• Dense Convolutional Network (DenseNet) and deconvolutional layers were integrated to improve feature reuse.
• Qualitative and quantitative experiments were conducted to validate the framework's effectiveness.

Main Results:

• The proposed framework demonstrated superior detection performance for medium and small stationary traffic objects.
• The system showed enhanced capabilities for nighttime traffic investigations, especially in complex intersection scenarios.
• Feature reuse optimization led to significant improvements in detection accuracy.

Conclusions:

• The developed framework effectively addresses the challenges of nighttime object detection in traffic scenarios.
• The integration of DenseNet and deconvolutional layers significantly boosts performance for critical object detection.
• This research offers a valuable tool for improving nighttime traffic safety and investigation efficiency.