Head and Shoulder Detection using Convolutional Networks and RGBD data
Objective
To develop a method for the real time detection of the heads and shoulders of humans in indoor environments using embedded systems.
Outcome
A complete system capable of high accuracy head and shoulder detection using RGBD data.
Challenges
- Run in real-time
- Low space and computation requirements to be able be ported to a dedicated image processing chip
- Perform reliably to be released as a consumer product
Core Model Architecture
A fast image processing algorithm for generating object proposals using the CHL algorithm, and a small convolutional neural network that processes the proposals and classifies them. For proposals predicted as belonging to humans, a head bounding box and shoulder keypoints are output.
Applications
This technology can be used in generating business analytics (number of customers, serivce time...), people counting, and congestion analysics.
Sample results
