Customer Demographic Analysis, for Sharp NEC

The Requirement:

We want to store the information of the demographics of customers that come to my store so that we can understand our customer base and also change arrangements in the store.

How did I solve it?

• Used the live CCTV footage coming into the network (passed to my colleague’s beautiful streamer module to chunk the stream to parts and process them), as input to a Computer vision pipeline
• The pipeline tracks customer’s faces using a Deep Learning based tracker, so that they are not repeated
• Predicts the age and gender of the customer using custom trained Deep Convolutional Neural Networks: Resnet18, Xception net on lots of public data and some methods from research papers
• Ran them in parallel and avoided repeative processing by skipping similar frames and running on those incoming during store open time only
• Gave options for different methods to use to get the best age and gender predictions for a person
• Options to run on CCTV stream, Raspberry Pi Camera, Webcam
• Reduced inference time using Post Training Quantization on these models
• Wrote all predictions to a Mongo Server
• Deployed using docker containers for easy deployment
• NOTE: Face data is not stored, just time, age and gender

Technologies:

• Container: Docker
• Programming Languages: Python
• Database: MongoDB
• ML Models: PyTorch (train), TensorflowLite (deployment), OpenCV

Team:

• Abhimanyu Bellam

Part of Guise AI Digital Signage, found here