Photo

Ruilin Liu


Ph.D. Candidate
Department of Computer Science
Rutgers, The State University of New Jersey
Office: CoRE 333
Email: rl475 [at] cs [dot] rutgers [dot] edu



About Me | News | Projects | Publications | Teaching | Resume

ParkScan: Parking Availablity Crowdsourcing based on Parking Decision Model

A main challenge faced by the state-of-art parking sensing systems is to infer the state of the spots not covered by participants’ parking/unparking events (called background availability) when the system penetration rate is limited. In this project, we tackle this problem by exploring an empirical phenomenon that ignoring a spot along a driver’s parking search trajectory is likely due to the unavailability. A typical parking search process can be modeled as a sequence of decisions. During a parking search trip, a driver frequently faces the decision moment to decide whether to take an available spot. This decision process is probabilistic and determined by the driver’s parking decision model where with the probability 1 - p(¬park|empty) .

An example of this parking process is illustrated in the following figure. A driver who is looking for a parking spot in a parking lot visited several spots, i.e., from spot 1 to spot 4. But since these spots were already taken, the driver continues to search until finding an available spot, i.e., spot 5. The driver evaluated this spot 5 by implicitly computing a probability, p(¬park|empty) based on the distance to his/her trip destination, i.e., the building on the right side of the figure. In this case, the probability p(¬park|empty) for ignoring spot 5 is not quite high, so the driver selects spot 5 and parks there.

The key idea of ParkScan is that if we can predict p(¬park|empty) for the spot along a driver's parking search trajectory, then we can use Bayesian rule to infer the availability probability based on the observation that the spot is not taken by the driver. Specifically, if p(park|empty) is approaching 1.0, passing-by without taking that spot implies that the spot is very likely to be unavaiable. The following three part are the main contribution of the project.

Parking Search Trajectory Dataset and Parking Decision Model

To study drivers’ parking decisions, we collect a large-scale dataset of parking trajectories in two public parking lots and developed the method to extract real parking decision events, i.e, either taking or ignoring an available spot during a parking search process. Using the features extracted from this valuable dataset, we build a data-driven model that reflects driver’s underlying decision strategy. To the best of our knowledge, this is the first dataset at scale that records real driver’s parking decisions with rich features. Based on this dataset, our model is the first data-driven parking decision model generated from large-scale real-world parking decision observations.

  • Dataset from Parking Lot 1
  • Dataset from Parking Lot 2
  • Realtime Parking Availability Estimation

    We design a crowdsourcing system to complement the state-of-art parking crowdsourcing systems particularly for their fine-grained background parking availability estimations. ParkScan takes parking searching trajectories from participant vehicles as input. By combining with historical knowledge and static parking facility layouts, ParkScan leverages a learned parking decision model to generate probabilistic parking availability estimations at the parking spot level.

    Large-scale Data Evaluation

    We evaluate ParkScan in both off-street scenarios (i.e., two public parking lots) and a city-scale on-street parking scenario. In the parking lot experiments, we utilize over 6,000 human parking search trajectories and real-world parking availability to conduct the evaluation (bottom-left figure). In the city-scale street parking scenario, we conduct data-driven evaluations using over 63,000 parking requests generated from real parking transactions in an urban area covering 35 blocks in Seattle, WA (bottom-right figure). Both of the experiments showed that with a 5% penetration rate, ParkScan reduces at least 12.9% of availability estimation errors for all the spots in the experimental scenario during the peak parking hours, compared to a state-of-the-art solution based on historical data. More importantly, even with a single participant driver, ParkScan cuts off 15% of estimation errors for the parking spots along drivers’ parking search trajectories.


    Ruilin Liu http://www.cs.rutgers.edu/~rl475
    Last updated on July 13, 2017