The embodiment of the invention provides a human flow statistics method, device and
system. The method comprises a step of acquiring continuous frames of images collected by an
image acquisition device, a step of inputting the continuous frames of images into a full
convolutional neural network obtained by training and generating a head confidence distribution map of each frame of image in the continuous frames of images, a step of determining at least one
human head detection target in each frame of image by using a preset target determination method for the head confidence distribution map of each frame of image, a step of obtaining a
feature matching result and motion smoothness of any
human head detection target in each frame of image, carrying out target association on the current frame and a previous frame of the current frame to obtain tracking targets according to the
feature matching result and the motion smoothness of any
human head detection target in each frame of image, distributing tracking identifiers to the tracking targets, and a step of counting the number of all tracking identifiers to obtain a result of human flow statistics. Through the method, the device and the
system, the accuracy and operational efficiency of the human flow statistics can be improved.