Features fusion with adaptive weights for pedestrian classification

In this paper, we study the problem of pedestrian classification, which could lead to an improvement of performance of the Pedestrian Detection Systems. Since the traditional approaches merely focus on the recognition of pedestrian, the device would keep alerting the drivers even if the pedestrians are walking on a safe track. We attempt to classify pedestrians in order to make those devices, equipped in the cars, more intelligent and pragmatic. We propose a method to extract features including HOGs (Histogram of Oriented Gradient), LTPs (Local Ternary Pattern), Color Names and to fuse them efficiently. The three features are weighted fused depending on the size of patches as well as each patch's gradient value which is computed via a 3*3 Sobel operator. Afterwards we will train a random forest with 50 discriminative decision trees, using the fused features. Our method is tested on the images of humans from INRIA dataset. The experimental results show that our method of features fusion, with adaptive weights assigned to the different features, yields a significant gain of 12.9% in mean AP (Average Precision) over the simple features concatenation framework. Accordingly, our method is practicable for classifying pedestrians.