Pedestrian Detection Based on Multi-Scale Fusion Features

The methods, integrating extra features into the features extracted from convolutional neural networks (CNNs) and using fusion features for pedestrian detection, have been considered effectively in recent years. However, the previous feature fusion methods only integrate the extra features with the final layer's detection features of CNNs. The main task of this paper is to prove the effectiveness of multi-scale fusion features for pedestrian detection. We propose a new network structure, which uses semantic segmentation feature as extra features and integrates it layer by layer into the feature pyramid structure of the detection network, in order to obtain multi-scale fusion features. Predictions are made on the fusion features independently. Experiments show that the method using fusion features in different scales achieves a better improvement in a relatively high speed.