Retail Behavioral Action Recognition Model

1. Project Title
Action Recognition Model for Detecting Customer Behavior in Retail Store

2. Overview
I need a deep learning-based action recognition model that can process video input from surveillance cameras and accurately detect customer interactions with store shelves. The model should classify each frame or short sequence into specific actions related to shopping behavior.

3. Target Action Classes
reach_to_shelf – User extends hand toward the shelf

picking_up – User retracts hand from shelf with an item

hand_in_shelf – Hand remains inside shelf (hovering or checking)

inspect_product – Customer inspects item in hand

inspect_shelf – Customer looks at shelf without touching

placing_back – Customer places an item back on the shelf

bold – No person/action visible in frame

4. Requirements
Train or fine-tune a video-based action recognition model (e.g., MSR-RNN, I3D, TSN, TSM, or custom CNN+RNN)

Accepts input video or live stream

Returns class label per frame or clip (windowed prediction)

Should be optimized for retail-like environment (indoor, crowded shelves)

Must include training pipeline + inference script

Accuracy benchmark preferred (confusion matrix, precision/recall)

5. Deliverables
Trained action recognition model (PyTorch/TensorFlow)

Annotated sample dataset (or guide to annotate videos)

Training & inference scripts

Exported model file

Documentation on:

Input format

Model architecture

How to run training & inference

FPS performance metrics

6. Tech Stack (Preferred)
Python

PyTorch or TensorFlow

OpenCV (for video I/O)

Optional: Keras, NumPy, Scikit-learn