Multi-Object Tracking based Roadside Parking Behavior Recognition

Abstract

Roadside parking spaces can alleviate the shortage of parking spaces, but there are some shortcomings to the charges for roadside parking. The popular charging methods at present mainly include manual charging, geomagnetic detection charging, meter charging, etc. These methods have certain limitations, such as high cost, difficult deployment, and low acceptance of people. To solve the shortcomings of roadside parking charges, this thesis proposes a scheme based on deep learning and image recognition. More specifically, the thesis proposes a scheme for detecting and tracking vehicles, recognizing license plates, recognizing vehicle parking behavior, and recording vehicle parking periods through the monocular camera to solve the problem of roadside parking charges. The scheme has the advantages of convenient deployment, low labor cost, high efficiency, and high accuracy. The main work of this thesis is as follows: 1. Based on the You Only Look Once (YOLO) algorithm, this thesis proposes a trapezoidal convolution algorithm to detect objects and improve the detection efficiency for the problem that the vehicle is far and small in the image. 2. Proposes a one-stage license plate recognition scheme based on YOLO, aiming to simplify the license plate recognition process. 3. Depending on the characteristics of the vehicle, this thesis proposes a feature extraction model of the vehicle, called the horizontal and vertical separation model, which use to combine with the deep Simple Online and Real-time Tracking (SORT) object tracking framework to track the vehicle and improve the tracking efficiency. 4. Uses a Long Short-Term Memory (LSTM) model to classify the behavior of the vehicle into three types: Park, leave, and no behavior. 5. Groups these modules together, and the engineering code is debugged a lot to realize a complete Roadside Parking Behavior Recognition (RPBR) system.