Previous methods found on camera geometry and projection matrix to select space image region for status classification. By utilizing suitable hand-crafted features, outdoor lighting variation and perspective distortion could be well handled. However, if also considering parking displacement, non-unified car size, and inter-object occlusion, we find the problem becomes more troublesome. To solve these issues in a systematic way, we proposed to design a deep convolutional network to overcome these challenges.

In this technology, we use three modules to achieve a robust system.
-First, we introduce a CNN-based deep network to extract more robust semantic features instead of relying on hand-crafted low-level features.
-Second, we integrate a STN into our deep network. The STN aims to adaptively crop, transform, and unify a 3-space input patch according to car sizes, occlusion patterns, and parking displacements.
-Third, in order to analytically solve inter-object occlusion problems, we group 3 neighboring spaces as an input unit. A multi-task loss function is designed to jointly consider the status estimation of the middle space and the occlusion patterns among neighboring spaces. A Siamese architecture is used to learn 3-space feature descriptor can preserve the “semantic” distances.

當前市面上，停車位管理大致以機械式與感測式的系統為主，這樣的系統在過去已經廣泛地被討論與應用，並得到相當良好且穩定的效果。然而，這樣的系統主要以感測器等硬體裝置為主，功能性上較不具備擴充的彈性，同時，系統維護上的成本也無法有效被降低，這也促使部分專家學者思考另外可行的改進方案。近年來，監控攝影機的使用越來越普及，相關的應用性逐漸地受到注目，因此已有不少研究將視訊監控系統應用於停車場管理上，希望透過軟體與演算法的開發，利用影像分析，自動偵測停車空位，以提供停車導引服務。