Pedestrian dangerous action recognition in infrared image based on Resnet18 network

Abstract

The effective identification of pedestrian dangerous actions at night was a core task of unmanned driving and intelligent assistant driving system. Limited by the network depth and learning ability of traditional convolutional neural network, the performance of the algorithm and its improvement were still unsatisfactory. Considering the imaging characteristics of the camera at night, this paper proposed an infrared pedestrian dangerous action recognition algorithm based on residual network to recognize pedestrian actions at night. Resnet18 network framework was adopted according to the characteristics of infrared images and the scale of problems. In order to adapt to the network input format, the infrared image in the database were preprocessed. The experimental results in the actual infrared pedestrian dangerous action dataset indicated that the mean precision of the proposed method for six types of dangerous actions was improved to 98.3%, and the average recall rate was improved to 98.1%, which was better than the traditional recognition method.