The invention discloses a method for detecting passenger congestion degree of urban rail transit based on convolution neural network. The method comprises the steps of processing the video to be detected, segmenting and extracting the motion residual image, taking the combination of original image and motion residual image as the input of convolution neural network algorithm, establishing a feature extraction block comprising at least one convolution layer and a maximum pooling layer, processing and calculating crowd state characteristics contained in the original image and the motion residualimage, and then combining the characteristics of crowd state with the characteristics of movement, building to contain at least one convolution layer, a feature fusion block of a maximum pooling layer and a full connection layer, subjecting to fusion processing, at the same time, constructing a classifier to train the convolutional neural network with the pre-fabricated training set with congestion label, so that the classifier can correctly detect the passenger congestion in the video to be measured, more comprehensive characterization of monitoring the passenger flow in the video, to achieve the detection of congestion, and improve the accuracy of the algorithm detection.