Download presentation
Presentation is loading. Please wait.
1
Airport Parking Space Navigation
Pingbo Pan, Ming Kong, Ting-Yao Hu, and Alex Hauptmann Language Technologies Institute, School of Computer Science, Carnegie Mellon University ABSTRACT Motivation: finding empty, convenient parking lot for user. Vacant Parking Space Navigation (VPSN) system detects vacant parking slots through camera and provide guidance to customers. VPSN system contains the following four modules: Parking Space Surveillance Vacant Space Detection Real-time Car Tracking Mobile Navigation App We setup cameras in Pittsburgh Airport parking area to acheive surveillance, collect data, and develop algorithms for for detection and tracking tasks. CAR TRACKING Locating the moving car in monitor screen in real time. Challenges: severe lighting/weather conditions object (car) angle changes a lot. real time process many distractors (cars in the parking slots) Approach: Tracking with classification given tracking result of frame, propose bounding box candidates. validate these candidates via classifiers Our algorithm performs well in accuracy, but not in efficiency. Next steps: consdering multiple cameras. PARKING SPACE SURVEILLANCE Incorporate with Pittsburgh Airport parking office. 5 ip cameras with different views setup around a section of long term parking area in Pittsburgh Airport A large scale video dataset for parking space surveillane collected ~8.8k hours (1.5 TB) Data in different weather, lighting conditions Potentially suitable for other event detection tasks VPSN System Vacant Space Detection Car Tracking Mobile Navigation App current location destination Parking Space Surveillance VACANT SPACE DETECTION VPSN system detect the available parking slots from surveillance image. Calibration: given the intrinsic/extrinsic parameters of camera, find out the parking slot regions. ~14k images annotated for experiment manual mapping used instead of automatic calibration formulate a binary classification problem Convolutional Neural Network (CNN) applied as classifier. In a preliminary experiment, we can achieve 0.95/0.94 precision/recall. . MOBILE NAVIGATION APP Turn-by-turn navigation service on mobile device Communicating with VSD and CT modules VSD module helps identifying the destination CT module report the current location of user’s car GPS signal enhance the instant report of current location adopting GPS signal only? Accuracy 3.5m at best (usually worse) hybrid approach Real-time navigation Android/iOS version Calibration & Segmentation Input Image Detection Result CNN classifier CONCLUSION AND FUTURE WORK Current progress shows the feasibility of VPSN system. Future works: improve the performance and efficiency of each modules full integration of all the functionality collecting data from real users SlidesMap Parking lot tool Why tracking with classification Turn left, and the empty slot will be on your right.
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.