Research on vehicle detection based on improved YOLOX_S

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vehicle detection in traffic scenes is challenged by small, distant targets and occlusions.
• Existing models like YOLOX_S can misdetect or omit these critical targets.

Purpose of the Study:

• To propose an improved YOLOX_S model for enhanced detection of small, long-distance vehicles.
• To address limitations in accuracy and feature representation for challenging traffic scenarios.

Main Methods:

• Model compression to optimize YOLOX_S inference speed.
• Integration of a coordinate attention module within a residual structure (Resunit_CA) to enhance small target feature attention.
• Addition of an adaptive feature fusion module to the PAFPN structure for richer feature extraction.
• Optimization of the decoupled head with Focal Loss to handle sample imbalance.

Main Results:

• The improved model achieved an average detection accuracy of 77.19% on the experimental dataset.
• Effectively alleviated issues of small-target missed detection and multi-target occlusion.
• Detection speed decreased to 29.73 fps, indicating room for real-time optimization.

Conclusions:

• The proposed YOLOX_S enhancement effectively improves the detection of small, distant vehicles in traffic.
• The modifications successfully reduce missed detections and improve handling of occluded targets.
• Further research is needed to balance accuracy gains with real-time detection speed requirements.