Identification of pedestrian attributes based on video sequence

Abstract

Analysis of pedestrian attributes in surveillance scenes is a challenging task for computer vision due to the influence of illumination and occlusion. Most existing algorithms focus on using static images to estimate results. However, these detectors often fail to appliance in videos. In this paper, We explore the GoogLeNet network that join Short-Term Memory (LSTM) to identify pedestrian attributes in continuous video sequences. Instead of dealing with a single frame, the network is used to predict the order images in run-on time. Extensive experiments are performed on video data in different monitoring conditions such as subway entrance, market and intersection. Results show that the method has achieved competitive performance on attribute recognition.