Human activity recognition with YOLOv8 and ResNet-50

Two-stage pipeline: YOLOv8 detects people, the box expands to capture context, a fine-tuned ResNet-50 classifies the activity. 96% test accuracy on Stanford40.

Status: Research
Period: 2025
Area: Deep learning
Code: Code public
Stack: Python
PyTorch Lightning
Ultralytics YOLOv8
ResNet-50
Stanford40

Test accuracy (Stanford40): 96%

Image
YOLOv8 person detection
Box expansion (context objects)
ResNet-50 classifier
Activity label

Fig. 2 Data flow, left to right (top to bottom on small screens).

Problem

Activities are defined by the person and the objects around them, so a tight person crop loses the signal.

Approach

A two-stage pipeline detects people with YOLOv8, expands the box to capture context objects, and classifies the activity with a fine-tuned ResNet-50 on Stanford40.

Results

96% test accuracy.