Robust wheelchair pedestrian detection using sparse representation

Detecting pedestrians with disability in surveillance videos is practical for the implementation of automated alert/assistance technology. This paper presents a novel approach for the dimensionality reduction which employs sparse representation to improve the generalization capability of a classifier. To characterize pedestrian with disability, we create directional maps by determining the dominant direction of motion in each local spatiotemporal region using 3D orientation filters, and then uses the maps in real-time surveillance settings to detect pre-defined types. Mathematically, the derived algorithm regards the input features as the dictionary in sparse representation, and selects the features that minimize the residual output error iteratively, thus the resulting features have a direct correspondence to the performance requirements of the given problem. Furthermore, the proposed algorithm can be regarded as a sparse classifier, which selects discriminative features and classifies the training data simultaneously. Experimental results obtained using the extensive dataset show the superior performance of our method and thus demonstrate its robustness with the novel sparse representation-based disabled pedestrian detector.

1. INTRODUCTION

Detecting pedestrians with disability in surveillance videos is practical for the implementation of automated alert or assistance technology. Specific applications include elevators that can respond automatically and the presence of the disabled, particularly in public transportation area, alerts staffs on duty to a need for assistance. Nevertheless, in the literature, few researches are proposed in pedestrian with disability detection. Huang et al. [1] detected wheelchairs with multiple views based on cascaded decision tree under a single-camera. They also imposed the tracking history to guide detection routes in the decision tree. However, using tracking history would make late decision. Furthermore, the video dataset used in [1] is captured with small viewing angles. In general video surveillance system of public area, the viewing angle would generally be larger, which makes the discrimination between wheelchair pedestrian and normal pedestrian a challenging problem since the observed shape of wheelchair would become insignificant. Therefore, how to discriminate wheelchair pedestrians from walking pedestrians is our main concern in this work. To detect normal pedestrian, though much progress has been made [2]–[6], recognizing pedestrians with high accuracy remains a challenging, unsolved problem. Significant difficulties of pedestrian detection in the surveillance setting include the facts that targets often have complex shapes due to their non-rigidness and can be of low-resolution because of the nature of a video camcorder. In [3], Viola and Jones proposed a boosted cascade for fast face detection. This kind of cascaded classifiers has been further extended in many other object detection problems, such as pedestrian detection [4], in which rectangles features are employed to construct the weak learners of the AdaBoost classifier for each stage of the cascade. Dalal and Triggs [5] proposed a people detection method for single images. The gradient-based features, histograms of oriented gradients (HOG), are developed to describe local gradient-orientation structure. In [4], Chen and Chen proposed a method that employed both intensity-based rectangle features and gradient-based ID features in the feature pool for weak-learner selection. The Real AdaBoost algorithm was used to select critical features from a combined feature set. Instead of using the standard boosted cascade, they employed both the stage-wise classification information and the inter-stage cross-reference information.