Revising deep learning methods in parking lot occupancy detection

TL;DR

This study extensively evaluate state-of-the-art parking lot occupancy detection algorithms, compare their prediction quality with the recently emerged vision transformers, and propose a new pipeline based on EfficientNet architecture.

Abstract

Parking guidance systems have recently become a popular trend as a part of the smart cities' paradigm of development. The crucial part of such systems is the algorithm allowing drivers to search for available parking lots across regions of interest. The classic approach to this task is based on the application of neural network classifiers to camera records. However, existing systems demonstrate a lack of generalization ability and appropriate testing regarding specific visual conditions. In this study, we extensively evaluate state-of-the-art parking lot occupancy detection algorithms, compare their prediction quality with the recently emerged vision transformers, and propose a new pipeline based on EfficientNet architecture. Performed computational experiments have demonstrated the performance increase in the case of our model, which was evaluated on 5 different datasets.

Figures (9)

• Figure 1: Image augmentation examples (simultaneously applied normalization, flipping, rotation, and noise injection). The figure demonstrates instances of empty (a) and occupied (b) parking lots, (c) shows the occlusion of the top part of a car and an empty parking lot.
• Figure 2: Examples of image-level JSON files for two possible types of annotations. Intersection-based annotations include 4 points per each parking lot defining the corresponding quadrangle, while patch-based annotations contain 2 points of a circumscribing rectangle.
• Figure 3: Demonstration of widgets' interfaces. (a) Position annotation widget enables a potential user to draw and erase quadrangles corresponding to parking lots as well as export the results in the JSON format. (b) The labelling widget requires pre-defined parking lot annotations and provides functionality to assign occupancy status to each entry.
• Figure 4: The final version of the EfficientNet-P model. The custom module follows the first seven blocks of the original EfficientNet-B0.
• Figure 5: Comparison of F1-scores for several considered models for various vision conditions. The bars are organized in ascending order with respect to the size of the categories; whiskers are based on the $Q1 - 1.5 \cdot \operatorname{IQR}$ and $Q3 + 1.5 \cdot \operatorname{IQR}$ values from 8 different splits.