Pedestrian Moving Direction Recognition using Ensemble of Convolutional Neural Networks

Abstract

Pedestrian moving direction recognition (PMDR) is a challenging task due to intra-class variation of pedestrians in cloths, their backgrounds and their walking poses. This paper focuses on the application of an ensemble classification approach based on Convolutional Neural Networks (CNN) for PMDR problem. The proposal method comprises four stages: (1) the pedestrian moving direction dataset generation, (2) the online data augmentation process, (3) the single CNN models and (4) the ensemble CNN models. Firstly, in order to obtain considerable number of features which fit the proposal problems, an original pedestrian dataset is produced. Afterwards, five single CNN models, as initial single strong classifiers, are built based on Cross-Validation. And the proposal online data augmentation is used during training process. Finally, the tuned single CNN models form ensemble CNN models by the proposal ensemble strategy and then, the best ensemble CNN model is selected for their performance evaluation. This paper highlights how ensemble learning technique like Averaging can dramatically improves the performance of a CNN based classification system. By correctly selecting the best combination strategy, the proposal method achieves better classification performance comparing with a widespread method like Random Forest + Histogram of Oriented Gradients (HOG), single CNN models, and several models trained based on some widespread CNN architecture.