Extending Parking Occupancy Detection Model for Night Lighting and Snowy Weather Conditions

Car parks today are usually monitored by camera systems in conjunction with the ground or thermal sensors. In the case, where the car parks are bigger and such options are not possible or would be very costly, new solutions are taken into consideration. One such solution is the use of an already in place commercial camera system, which would be enhanced by neural networks for parking space classification in real-time. Some of the more popular works in this area are the PKLot and CNRPark papers, which include the generation of datasets for visual occupancy detection of parking lots for the training of neural networks. These datasets contain images in varying weather and lighting conditions, such as sunny, overcast, or rainy, including partial occlusion due to obstacles such as trees, other cars, or lampposts. In this work, we attempt to create yet another dataset that would capture parking lots in night lighting conditions and snowy weather conditions. In the CNRPark paper, a CNN model called mAlexNet has been trained, which is a minimalist version of the AlexNet. This lightweight model is suitable for real-time car parking occupancy detection for smart camera devices with low computational power. In our work, we attempt to use transfer learning to further improve the robustness of this model using our newly generated dataset. Our experiments will be conducted using a Raspberry Pi model 4B with 4GB of RAM equipped with a Raspberry Pi High Definition Camera in conjunction with the Raspberry Pi Wide Angle Lens.