Abstract: The disclosed technology for building an edge convolutional neural network (CNN) system for IoT includes training on-site processors to analyze image data and identify motorized vehicles, bicycles and people in near-real-time, using a big cloud CNN, a small cloud CNN and an on-site CNN. At least five hundred site-specific images from cameras are analyzed using the big cloud CNN to produce a machine-generated training set that includes an image that has regions, and for each region, coordinates of bounding boxes for objects in the region, and classification of contents of the bounding boxes as a motorized vehicle, bicycle or person; and the training set. The machine-generated training set gets used to train the small cloud CNN; and coefficients from the trained small cloud CNN get transferred to the on-site CNN, thereby configuring the on-site CNN to recognize motorized vehicles, bicycles and people in images from the cameras in near-real-time.

Abstract: The disclosed technology for building an edge convolutional neural network (CNN) system for IoT includes training on-site processors to analyze image data and identify motorized vehicles, bicycles and people in near-real-time, using a big cloud CNN, a small cloud CNN and an on-site CNN. At least five hundred site-specific images from cameras are analyzed using the big cloud CNN to produce a machine-generated training set that includes an image that has regions, and for each region, coordinates of bounding boxes for objects in the region, and classification of contents of the bounding boxes as a motorized vehicle, bicycle or person; and the training set. The machine-generated training set gets used to train the small cloud CNN; and coefficients from the trained small cloud CNN get transferred to the on-site CNN, thereby configuring the on-site CNN to recognize motorized vehicles, bicycles and people in images from the cameras in near-real-time.