Research Dossier / 02

Human Body Parts Detection using Improved YOLOv5 with Multi-Layer Attention Network (MLA-NET)

Designed an enhanced YOLOv5-based detection framework integrated with a Multi-Layer Attention Network (MLA-NET) for accurate human body part detection in disaster rescue scenarios.

Short LabelMLA-NET YOLOv5
Institution
Research under Dr. Soumajit Pramanik, IIT Bhilai
Role
Computer Vision Researcher
Active Period
2023 - Present
YOLOv5
PyTorch
OpenPose
Real-ESRGAN
Deep Learning

Method + iterations

Research steps, prototypes, and refinements along the way.

  1. Developed a multi

    layer attention module combining channel attention and pixel attention to improve feature refinement.

  2. Automated ground

    truth generation using OpenPose — based body joint estimation.

  3. Enhanced low

    resolution disaster footage using Real — ESRGAN super — resolution models.

  4. Implemented frame extraction pipelines for video

    to — image dataset creation.

  5. Built augmentation sets targeting smoke/noise/blur artifacts common in rescue operations.

  6. Achieved improved robustness in detecting partially occluded or irregularly posed victims.

Visual Notes

ATTENTION ARCHITECTURE

ATTENTION ARCHITECTURE

Channel + Pixel attention integration improving detection accuracy in complex rescue environments.

TRAINING PIPELINE

TRAINING PIPELINE

Dataset generation workflow from raw video streams to curated frame-level labels and model-ready samples.

INFERENCE UNDER STRESS

INFERENCE UNDER STRESS

Validation runs across low-light, occluded, and compressed inputs to stress-test deployment resilience.

← Back to archiveNext / Termx