A People Counting Method Combining Convolutional Neural Networks and Trajectory Prediction

Abstract

The invention relates to a method for counting people by combining convolutional neural network and trajectory prediction, comprising the following steps: using frame difference method to segment crowd clumps contained in video; distinguishing sparse crowd clumps from dense crowd clumps; sparse crowd clumps Cluster, the HSV color space obtained by transforming the RGB image by the formula, using the selection search algorithm to pre-determine the pedestrian position in two different color spaces, and merging and removing the repeated areas of the two spaces to obtain the pedestrian area position; using convolutional neural network The network extracts features, and selects the grid loss function Grid Loss to train the network in blocks to realize the recognition of the local position including the face and torso of the occluded pedestrian; for dense crowds, extract the characteristics of the crowd density distribution map and establish a multivariate Regression model and estimate the number of people; for occluded pedestrians, use Markov model chain to predict their walking trajectory, then lock their position and count the pedestrians.

Description

technical field [0001] The present invention relates to the field of target recognition, in particular to a method for counting people in combination with convolutional neural network and trajectory prediction. The method is mainly used in video people counting represented by deep learning and trajectory prediction for target detection. Background technique [0002] People counting is a research direction with practical significance in the field of intelligent video surveillance in recent years. It is mainly divided into people counting methods based on target detection and people counting methods based on feature regression. Both types of methods use supervised machine learning. In addition, there are tracking trajectory clustering methods based on unsupervised learning. In supervised learning methods, pedestrian detection based on the HOG algorithm is one of the widely used methods. This method constitutes pedestrian features by calculating and counting the gradient direct...

AI Technical Summary

Problems solved by technology

Although these algorithms have extracted effective and high-quality pedestrian features and established accurate prediction models, they still cannot solve the impact of factors such as occlusion, illumination changes, and uneven crowd distribution on detection.

Images