TOP  >  Thesis/Dissertation  >  Inter-Camera Pedestrian Matching by Feature Selection Based on the Confidence Metric

Inter-Camera Pedestrian Matching by Feature Selection Based on the Confidence Metric

It is useful to analyze trajectories of people in commercial spaces such as shopping malls. For example, by analyzing the tendency of people's trajectories, we can investigate the arrangement of the shops in a mall.

There are many approaches to automatically obtain the people's trajectories. The approach which uses cameras is suitable for our case, because it can be used at indoor scenes and there have been a lot of surveillance cameras installed in commercial spaces which can be used.

To obtain people's trajectories from images of surveillance cameras, we have to match persons in images of different cameras. For the matching, we use the similarity between extracted features of the persons, such as color and designs of their clothes. Each type of features has parameters, which can control how strict the feature comparison is and how detailed each feature description is.

By using color as a feature, two different persons can be distinguished based on differences in their clothes color. However, a person cannot be distinguished from others wearing clothes of similar color. In this case, if we use the design of their clothes as features, we may distinguish them. On the other hand, when the orientations of two persons are different between camera' views, it is difficult to match them by using the design of their clothes, indeed the clothes design may not be visible among views. In this case, we might be able to match them by using the color of their clothes as features, which are robust to the change of orientation. This illustrates the need to use appropriate type of features and parameters for each person observed by cameras and each observation environment. However, it is difficult to know which type of features and which parameters we should use beforehand.

We propose a method to automatically select the bests feature ans parameter for matching a specific person in a specific observation environment. We use a confidence metric to show how different the similarities are between the most similar pair of a target person and the other pairs of the target person. Using the inappropriate type of features and parameters, gives similar value for all possible match. In this case, the confidence value may be low and there is a low probability that the two persons whose features are the most similar are the same person. On the other hand, when we use the appropriate type of features and parameters, the features of the target person tends to be similar only with the features of the same person. Then, the confidence value will be high and there is a high probability that the two persons whose features are the most similar are the same person. For each type of features and set of parameter values, we evaluate the confidence metric. We selected the type of features and parameters of which confidence value is the highest and using them we evaluate the similarity between persons. In this way, we can use appropriate type of features and parameters for each person observed by cameras and each observation environment.

We confirmed the effectiveness of our method by performing some experiments on videos obtained from two different cameras in a mall. We extracted Color Histogram and SIFT as features, and for each features we used two different parameter values, so we get four possible different pairs. Using our method, we could correctly match eighteen persons out of twenty. While we could correctly match seven to fifteen persons by applying identical pair for each person. These results show that our method can select appropriate type of features and parameters and improves the correct matching ratio.

Our future works aims to improve the matching results by using other types of features such as facial features. We have to evaluate our proposed method with many types of features and confirm its effectiveness in different environments.