New End-to-End Strategy Based on DeepLabv3+ Semantic Segmentation for Human Head Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Semantic Segmentation

Background:

• Object detection in computer vision identifies objects and their positions in images.
• Key applications include safety systems, control systems, and particularly head/person detection for road safety and surveillance.
• Current methods require continuous improvement for accuracy and efficiency.

Purpose of the Study:

• To develop a novel, high-performance person detection system.
• To enhance semantic segmentation accuracy for computer vision tasks.
• To evaluate the proposed approach against state-of-the-art models.

Main Methods:

• A new approach utilizing two parallel DeepLabv3+ models was developed.
• A semantic segmentation model was implemented using a methodology with two types of ground truths derived from bounding boxes.
• The approach was tested on private and public datasets.

Main Results:

• The proposed system achieved a global accuracy of 99.14%.
• Comparative analysis showed superior performance over SegNet and U-Net models.
• The strategy proved efficient for semantic segmentation tasks.

Conclusions:

• The developed strategy offers an efficient deep neural network model for semantic segmentation.
• This approach is effective for human head detection and applicable to broader semantic segmentation challenges.
• The method demonstrates significant potential for improving computer vision safety and surveillance systems.