How AI Predicts Real-Time Parking Availability Without Sensors

Background

Parking is a basic need for drivers, yet finding a parking space remains a persistent pain point, especially in dense office districts, commercial areas, and busy hospitals. The time spent searching for parking not only wastes drivers' time but also aggravates traffic congestion, reducing overall city traffic efficiency.

Some readers may wonder whether knowing in advance if a destination has convenient parking could help them decide whether to drive or where to park before arrival. Existing solutions that provide online parking information rely on hardware devices, which are costly, have high operational expenses, and suffer from accuracy issues due to interference from other vehicles, resulting in limited coverage. With the rise of artificial intelligence, Baidu Maps has applied AI technology to predict parking space availability for venues lacking real‑time data, a result published at the KDD conference.

The key question is how to use AI to predict real‑time parking availability for sensor‑less lots.

Overall Idea

The insight is that parking demand is fundamentally driven by the human flow around a parking lot’s POI. Therefore, the modeling focuses on capturing the spatio‑temporal characteristics of the POI’s foot traffic.

A DNN‑based method is proposed, consisting of three parts that model temporal proximity, periodicity, and static environment. Feature engineering leverages Baidu Maps’ rich crowdsourced data (e.g., location heat, navigation heat, POI search heat) to characterize parking demand. The approach is applicable to other spatio‑temporal parking problems and has been deployed at scale in Baidu Maps.

Technical Solution

The system comprises three components: grid computation, real‑time feature calculation, and online prediction service.

a) Grid Computation

Using an offline distributed computing framework, it performs three tasks: feature extraction, generation of parking‑availability samples, and model training.

Static features: POI type, search heat, visit heat, construction year, etc.

Sample processing: collect real‑time parking data from a few equipped lots, label the data, and store it persistently.

Model training: extract static and dynamic features from historical data, train a separate model for each POI type to improve overall prediction accuracy.

b) Real‑Time Feature Calculation

Through a streaming architecture, it computes real‑time navigation, search, and location features to capture the dynamic flow of people around each parking lot.

c) Online Prediction Service

It fuses POI‑related features, map‑search features, real‑time geographic and navigation trajectory information into a feature vector that predicts parking difficulty.

Feature Engineering

Four categories of features are constructed: real‑time crowd features, POI features, weather features, and temporal features.

Real‑Time Crowd Features

These include location heat and navigation heat.

Location feature: the real‑time location heat of the POI, reflecting how human movement patterns can indicate vehicle movement.

Navigation feature: the heat of navigation queries entering and exiting the parking lot, representing parking demand.

Analysis shows a strong correlation between these crowd features and the number of available parking spaces for both office buildings and shopping malls.

Weather Features

Weather and wind speed affect traffic flow; adverse conditions (low temperature, rain, snow, strong wind) typically reduce vehicle volume and increase parking availability.

POI‑Related Features

Parking availability is closely tied to the type of the surrounding building. Seven parking‑lot categories (apartment, office, mall, restaurant, hospital, park, entertainment) are identified, and the building’s construction year is used to infer total capacity. Different types exhibit distinct daily patterns.

Parking Availability Prediction Model

The model uses three separate networks, following a wide‑and‑deep paradigm:

A temporal LSTM network extracts periodic features from the input sequence.

A second LSTM network captures proximity features.

A multilayer perceptron processes static POI attributes.

The outputs of the three networks are fused and fed into a final classifier.

Features for each time slice include spatial movement information, POI data, weather data, holiday indicators, and the current timestamp. Periodic features are derived vertically along the time series, while proximity features are derived horizontally by looking back a fixed number of slices.

Further Work

Future directions include extending real‑time prediction to future time horizons, employing newer deep‑learning models to better capture dynamic crowd, traffic, and spatial topology information, and incorporating graph convolutional networks to leverage spatial relationships among parking lots for both current and future predictions.

References

[1] Yuecheng Rong, Zhimian Xu, Ruibo Yan, and Xu Ma. 2018. Du‑Parking: Spatio‑Temporal Big Data Tells You Realtime Parking Availability. https://doi.org/10.1145/3219819.3219876

[2] Weijia Zhang, Hao Liu, Yanchi Liu, Jingbo Zhou, and Hui Xiong. 2020. Semi‑Supervised Hierarchical Recurrent Graph Neural Network for City‑Wide Parking Availability Prediction. Proceedings of the AAAI Conference on Artificial Intelligence 34(04): 1186–1193. https://doi.org/10.1609/aaai.v34i01.5471