Vehicle and pedestrian video counting method and device based on neural network

Abstract

The invention discloses a vehicle and pedestrian video counting method and device based on a neural network, and belongs to the field of video counting. The method comprises the following steps: acquiring a current frame from a video; segmenting the current frame into a plurality of rectangular grids by adopting a YOLO algorithm; detecting a target object of which the central point is located in each grid; judging whether the detected target object is a detected category or not, and if not, determining that the target object is a new category; performing feature extraction and prediction on the detected target object; if predicting that a target object is a vehicle or a pedestrian, matching and tracking the features of the target object extracted from the previous frame and the current frame; and counting vehicles or pedestrians by detecting whether the target object in the current frame passes through a specified counting line or not. The device comprises an acquisition module, a segmentation module, a detection module, a classification module, a prediction module, a tracking module and a counting module. Errors between large-scale target counting and real data after training aresmall.

Description

technical field [0001] The present application relates to the field of video counting, in particular to a neural network-based method and device for video counting of vehicles and pedestrians. Background technique [0002] Counting the number of pedestrians through digital video has a certain research basis. Some of the traditional methods have been considered classic and have many commercial and academic applications. Most of the algorithms for pedestrian counting come from the segmentation of foreground and background in the working scene, and the foreground objects are determined by contour analysis or blob detection threshold. [0003] In recent years, with the development of deep learning, deep learning has made many breakthroughs in the field of computer vision due to its high-performance image classification and target detection and huge computing power. In particular, in 2016, Joseph Redmon and others proposed the YOLO algorithm, which demonstrated a new idea that ...

AI Technical Summary

Problems solved by technology

[0004] However, there are some problems in the classic method based on separating the foreground. First, the shadow of the object is difficult to separate from the object itself, and it is difficult to distinguish, and this method does not provide a direct classification method to detect objects, and is affected by scene lighting and image noise. bigger

Most deep learning methods improve the accuracy of the video detection process, but the detection speed is slow, and some do not have the ability to classify

Images