Method and system of real-timely estimating dimension of signboards of road-side shops

A system, comprising:

• an edge device, a photographing device that are onboard a first moving vehicle, wherein the edge device comprises a GPU (Graphics Processing Unit), a processor, one or more computer-readable memories and one or more computer-readable, tangible storage devices, a GPS (Global Positioning System) Unit, a GSM (Global System for Mobile Communications) module, a power management unit, a supervisor unit, and a control unit, and a portable computing toolkit,
• wherein the photographing device is connected to the edge device for captured video stream of roadside scenes to be channeled to the edge device's video buffer, and the photographing device is placed inside the first moving vehicle on the windshield or on the top of the first moving vehicle, to perform operations,
• the operations comprising:
• capturing and storing a first set of still images or video footages that contain a first set of signages by using the photographing device,
• extracting a first set of spatial features from the first set of still images or video footages, detecting a first set of signages and a first set of masks by using a Fully Convolutional Neural Networks (FCNN) based on the first set of spatial features, wherein the step of detecting the first set of signages and the first set of masks comprises calculating a mask loss value using Binary Cross Entropy (BCE) for each of the first set of masks, and calculating a first total loss value based on the mask loss value for each of the first set of masks, and using the first total loss value to adjust, using a stochastic gradient descent, a plurality of weights associated with the FCNN, via back-propagation, to steer the FCNN to learn from the first set of still images or video footages,
• localizing the first set of signages by using a tracking system to identify ownership information of each sign of the first set of signages to produce a first set of identification associated with the first set of signages,
• passing the first set of signages and the first set of masks into a pre-trained Multi-Layer Perceptron (MLP) neural network to estimate a first set of dimension values by regressing dimension values based on a loss function applied on the first set of masks, wherein each dimension value in the first set dimension values contains a height value and a width value, and recording on the edge device, and
• transmitting and registering the first set of signages in association with the first set of identification and a first set of dimension values to a remote database,
• wherein the extracting, detecting, and passing steps are carried on by one or more computer programs loaded in the one or more computer-readable memories and executed by the processor or the GPU.