Method for Identifying Still Objects from Video

A non-transitory computer program contained in a non-transitory storage medium comprising program code instructions which executes a method of detecting a stationary object from a static video by a computer hardware device by use of inference information of Deep Learning object detector, the method comprising:

• obtaining inference result from the inference of Deep Learning object detector on the static video for each of a plurality of predetermined scales, wherein the inference result has the form of a predetermined multidimensional array which corresponds to a plurality of predetermined channel configuration;
• temporarily storing the inference results in a buffer during a predetermined frame period;
• calculating an average inference value for each of the plurality of scales by performing an average operation for each of the plurality of scales on attribute values of the plurality of inference results which are temporarily stored in the buffer, wherein the attribute values correspond to probability of object;
• acquiring data of one or more bounding boxes from the static video by performing bounding box decoding on the inference result of Deep Learning object detector;
• obtaining a stationary score for each of the bounding boxes by comparing the one or more bounding boxes obtained by the bounding box decoding with the average inference value calculated for each of the plurality of frames;
• and
• setting a bounding box having the stationary score greater than a predetermined threshold value as a stationary object.