Research On Indoor Pedestrian Detection Based On Visual Apparentness And Integral Channel Feature

In recent years, computer vision has begun to emerge in the field of intelligent house, and object detection and recognition is the research focus in the field of Computer Vision. How to make the pedestrians detection more accurate under the indoor environment is especially important to the smart system of home. Aiming at the problem of high false detection rate of pedestrian under indoor environment, in this thesis, an indoor pedestrian detection algorithm based on saliency theoretical was proposed. With this method, the false detection rate has been effective reduced to a controllable range, and the pedestrian detection algorithm has been improved to get a shorten detection time.First, several common used pedestrian detection algorithms based on statistical learning which was state of the art were discussed in this paper, including HOG (histogram of gradient) combine with Support Vector Machine algorithm, deformable partial model algorithm and multi-feature fusion algorithm (integral channel feature). What’s more, their advantages and disadvantages were presented respectively. Due to pedestrian detection datasets mostly established for pedestrian detection in common scenarios, and lacked of indoor pedestrian samples. We re-established a relatively well collected indoor pedestrian datasets. In order to improve the overall speed, the binarized normed gradients (BING) algorithms based on saliency theory which could both ensure the accuracy of detection and reduce the number of candidates windows to be detected was adopted.Secondly, analyzing the selection mechanism of negative sample in the Adaboost classifier’s model-phase, we found that the generation of random negative samples caused the high false drop rate. For improving the high false alarm, the candidates windows generate algorithm based on saliency (BING) also was introduced. In this way, negative samples were selected by which was more like an ’Object’ or easy confused with the pedestrian. So that, more stable and higher distinguish classifier model could be trained. Aiming to the specificity of indoor environment, we took the head and neck part of pedestrian as positive samples which could effectively prevent the occurrence of partial occlusion. Furthermore, establishment of training samples for the indoor environment could make the trained classifier model to get a better detect performance. Experimental results showed that this method used in indoor pedestrian detection could get a good performance of robustness and effectiveness.