PKLot – A robust dataset for parking lot classification

• We present a robust image dataset for parking space classification.
• We evaluate textural-based descriptors for parking space detection.
• Classifiers are evaluated on different parking lots.
• Classifiers are evaluated on parking lots that were used for training.

Outdoor parking lot vacancy detection systems have attracted a great deal of attention in the last decade due the large number of practical applications. However, a common problem that researchers in this field very often face is the lack of a representative dataset to perform their experiments. To mitigate this difficulty, in this paper we introduce a new parking lot dataset composed of 695,899 images captured from two parking lots with three different camera views. The acquisition protocol allows obtaining static images showing illumination variance related to sunny, overcast and rainy days. We believe that researchers will find this dataset a very useful tool since it allows future benchmarking and evaluation. The dataset is currently available for research purposes upon request. To gain a better insight into this dataset we have evaluated two textural descriptors, Local Binary Patterns and Local Phase Quantization, with a Support Vector Machine classifier to detect parking lot vacancy. In the experiments where the same view was used for both training and testing, we have reached outstanding recognition rates, greater than 99%. The main challenge, though, lies in building a general classifier that is able to detect parking spaces from the parking lots that were not used for training. In this sense, the best result achieved by the texture-based classifier was about 89%. The observed drop in terms of performance shows that additional investigation is necessary to create classification schemes less dependent on the training set. Other researchers can use these results as a baseline performance when testing their own algorithms on this dataset.