Pedestrian Localization Algorithm Based On Feature Selection

Object localization is an important topic in computer vision,and pedestrian localization is the most common issue of object localization problem.It is extensively studied in the field of intelligent security,automated driving,human-machine interaction and so on.Pedestrian localization methods can be divided into pedestrian tracking method and pedestrian detection method.Although the existing pedestrian localization algorithms have achieved good performance,however,it is still a challenging task to achieve robust pedestrian localization in complex scenes.Firstly,most of the existing tracking-by-detection algorithms do not utilize features specifically tailored for pedestrians to establish the appearance model.Consequently,these tracking algorithms cannot accurately describe the inter-class distance and intra-class distance of pedestrian targets.Secondly,for the pedestrian detection task,there is still a semantic gap between the detection results and human visual judgments due to the existence of various scales,pose changes,occlusions and so on.In this paper,we provide an overview of object localization methods,including object detection methods and object tracking methods.To tackle the abovementioned problems of the existing problems,we propose three innovative pedestrian localization methods from the perspective of feature selection.Firstly,we use the selected hand-crafted features to improve the classic Tracking-Learning-Detection framework.The ACF and LOMO features,which can respectively provide good measures for the inter-class distance and intra-class distance of pedestrians,are selected to establish a robust pedestrian appearance model.Experimental results show that our appearance model can improve the tracking success rate by 3%.Secondly,we propose to achieve pedestrian detection using the combination of hand-crafted feature and neural network feature.The framework of the proposed algorithm is cascading the ACF proposals and Fast RCNN.To be more specific,we improve the training strategy of ACF algorithm and use online hard example mining strategy in training Fast RCNN model.As a result,the ACF proposals can obtain a higher recall rate and the whole detection algorithm is able to achieve 17% miss rate on Caltech pedestrian dataset.Thirdly,we fully utilize CNN features to perform pedestrian detection task,and train an end-to-end multi-scale convolutional neural network.We propose to extract multi-scale context features and hierarchical convolutional features for target representation.Our network can effectively solve the problem caused by the scale variation of pedestrian targets.The proposed algorithm can obtain 9.51% miss rate on Caltech pedestrian dataset,which is 0.4% better than original multi-scale convolutional neural network algorithm.We evaluate all the proposed pedestrian localization methods on the Caltech pedestrian database.Experimental results show that the proposed pedestrian localization algorithms can effectively improve the success rate and accuracy of localization results.