City Scale Monitoring of On-Street Parking Violations with StreetHAWK

Abstract

Unauthorized parking on city streets is a major contributor to traffic congestion and road accidents in developing nations. Due to the large scale and density of this problem, citywide (manual) monitoring of parking violations has not been effective with existing practices. To this end, we present StreetHAWK: an edge-centric, automated, real-time, privacy-preserving system; which leverages the rear camera of a dashboard mounted smartphone, and performs visual scene and location analytics to identify potential parking violations. We realize this system by overcoming the challenges of: (i) small object identification in various non-standard setups by extensive training on a deep learning based convolution detection model; (ii) limited violation assessment range of 15 m (a constraint of the phone's single camera unit) by augmenting it with a short-term historian and GPS for meeting the 100 m measurement violation guideline; and (iii) erroneous mobile scene analysis instances by lightweight filtering techniques that piggyback on the mobility of the camera and multi-modal sensing clues. The evaluation results obtained from real-world datasets show that StreetHAWK: (i) has three times higher accuracy in identifying small sized objects than other competing embedded detectors; and (ii) localizes these objects from a moving vehicle with a worst-case error of less than 5 m. On-the-road experiments show that StreetHAWK, running at a speed of 5 frames per second (FPS) on a typical Android smartphone, was able to detect (on an average) 80% of the parking violations.