Inference Models

Gravio can infer camera images using TensorFlow , a Google system used for Artificial Intelligence and Machine Learning. Two types of TensorFlow files are available:

• TensorFlow and
• TensorFlowLite.

There is a difference in the number of deep learning layers in the inference process:

• TensorFlow has 32 layers and
• TensorFlowLite has 16 layers

While TensorFlow is more accurate in inference, it requires more processing time and memory. We suggest that you verify with your environmental conditions such as lighting, contrast, camera resolution etc. if the TensorFlowLite model is sufficient.

Available Inference Files

The following inference files can be used in Gravio per default:

Recognition of congestion

Recognizes the degree of congestion in the area (range) captured by the camera.
16Feet, 32Feet, and 48Feet are approximate distances to the area recognized by the camera.
48Feet is a wider range than 16Feet, so please use the larger number to identify people in the distance.

Outputs: Integer of number of people detected (counting heads, optimised for an indoor setting.)

Number of people counting

NumberOfPeopleTensorFlow

This inference counts the number of people recognized by the camera.

Outputs: Integer of number of people detected (counting full bodies, optimised for an outdoor setting.)

Weather recognition

WeatherTensorFlow

Recognizes the weather in the sky recognized by the camera.

Outputs: windy, thunder, cloudy, typhoon, snowy, sunny, rainy, unknown

If you have a free or basic subscription, you get the people counter and weather models in the bundle. If you have the Enterprise subscription, you can upload your own AI/ML models to your Coordinator back end and push them to the Gravio HubKits. Note, the models are general purpose models and they may not work specifically for your circumstance. If you need more accurate models, we recommend training them specifically for your setting.