A multiclass boosting approach for integrating weak classifiers in parking space detection

Recently, Huang's method [1] has proposed to use a 3D parking space representation for parking space detection. Following a generative process, the approach treats a parking lot as the collection of many parking spaces. Each space is modeled by a 3D cube. Each 3D cube is composed of multiple 3D surfaces. If projecting those 3D surfaces onto the image, many image patches of a parallelogram shape would be determined; each patch may reveal some weak information that could be used to infer the parking status. In order to transfer the image feature into status information, the approach trained a weak classifier for each image patch. Finally, by combining these weak classifiers, this approach could well determine the parking status. However, we found that the system weights for combining the weak classifiers in Huang's method are manually selected. This might not be suitable since different classifiers usually have different class discriminative ability. To address the issue, we proposed a multiclass boosting method to incorporate these weak classifiers through a back-propagation learning process.