University of Pittsburgh researchers have developed a novel system to detect vacant on-street parking spots. This passive mobile sensing solution requires no specialist hardware and is smartphone-based using a deep neural network (DNN) to analyze sounds reflected from the roadside and predict parking space availability. This novel system could be part of a crowdsensing system, where vehicles can passively detect on-street parking availability. Information can be provided to a phone application allowing users to identify available parking locations, saving time and fuel while reducing carbon emissions.
Description
Finding on-street parking is a great cause of driver frustration in many busy cities. In the city of Los Angeles alone, each year over 17 million gallons of fuel and 47,000 hours are wasted by drivers looking for a parking spot while circling the streets. There exists a technology gap between static sensors which are costly for comprehensive coverage and mobile sensing which requires specialist software installed in vehicles. This novel system has been designed to detect and analyze environmental sound reflections while a vehicle is driving to predict the availability of on-street parking. Information can be shared with others to assist drivers, reducing the need to waste time and fuel driving around a city in the hunt for a parking space.
Applications
• Smartphone application to find on-street parking
Advantages
Current on-street parking sensors are either static or mobile. Static sensors are usually limited to monitoring single parking spaces with multiple sensors required for comprehensive coverage, which increases costs. While mobile sensing is used in some cities, it requires specialist hardware installed in vehicles and is not widely used.
The novel system developed here only requires a smartphone to collect sounds from the vehicle. Audio is collected through a smartphone microphone, pre-processed for data segmentation, and a DNN model used to predict parking space availability. No specialist hardware is needed, and data can be collected passively without any action from the user. This system could be easily accessible to many users resulting in large amounts of information on parking space availability being collected for users in search of on-street parking.
Invention Readiness
Using an electric and a traditional electric vehicle fitted with different types of smartphones and cameras, both audio and visual data were collected to train the DNN model. Aerodynamic noises in the 8 kHz and 17 kHz regions were used for the training of the model and changes in reflected noise was compared to the ground truth (i.e., camera images) to predict parking space availability. Testing of the prototype system confirmed 93.51% prediction accuracy and 96.55% true positive rate. Work is now required to optimize this system to include considerations around interference from traffic and weather noise, as well as speed and car model variables.
