Traffic Density Estimation for Traffic Management Applications Using Neural Networks

1. Introduction

Due to the step-up in the number of vehicles everyday, traffic over-crowding and jams have become quite common. Recently with the technological advancement in Intelligent Transport System (ITS), the transportation system authority can acquire information used in traffic engineering such as number of vehicles and speed of the vehicles (Takayuki, 2017). Traffic jams not only affect the routine lives of human but also lead to a rise in the cost of transportation. This makes developing an automated management system for traffic unavoidable. Traffic density estimation is the most important task that should be done for ITS. Traffic density is one of the elemental dealing’s variables used in molding road traffic kinetics. It quantifies the number of vehicles is on route. It is the first harmonic building block for many traffic management applications like advising best itinerary to user based on stream traffic conditions, live updates about congestion and traffic jams etc. Thus, the main aim of this work is on finding the density of traffic in a day-to-day scenario (Zhiming, 2018). To obtain information on the routine of vehicles and their speed through CCTV, the primary thing to be done is detection of vehicles. There are different methods that can be used for detection of vehicles. The application of image processing and computer vision techniques in detection of vehicles improves improves these strategies of traffic information assortment and road traffic observance (Girshick, 2014, Chen, 2007).

Traffic density can be estimated in three ways and are as follows:

• 1.

  Sensor based method: With the technological advances in microchip, RFID (Radio Frequency Identification) and inexpensive intelligent beacon sensing technologies, sensing systems for ITS have been extensively developed (Karandeep, 2017):

  • a.

    Inductive Loop Detector: A prominent sensing technology currently being used for vehicle detection is induction loop. An induction or inductive loop is an electromagnetic communication or detection organization. It makes use of an oscillating magnet or moving current to stream electric current in a wire situated nearby. These loops can be placed below the bed of the road to find vehicles as they cross the loop's magnetic flux. The simple loop detectors just count the number of vehicles throughout a unit of your time that skip the closed circuit, whereas additional refined sensors estimate the appraisal of speed, category of vehicles and also the space between them. Loops may be placed in a single lane or across multiple lanes, and that they work with vehicles moving at variable speed (Karandeep, 2017).

  • b.

    Mobile Sensor: Inductive sensors discussed above have limitations related to their installation, data collection etc. Using vehicles as sensors can remove these limitations as sensors are put in and maintained easily on vehicles than on the road and thus the traffic information is collected from everyplace (Zhiming, 2018).

• 2.

  GPS based Methods: Smartphones are now equipped with many sophisticated built-in sensors such as Global Positioning System (GPS) receivers, accelerometers, gyroscopes, cameras, and microphones. These sensors can be exploited to sense the traffic data. Without any new equipment, a person with a smartphone can turn any vehicle into a mobile traffic sensor (Takayuki, 2017).

• 3.

  Computer Vision based method: Here, camera feed from surveillance cameras placed at road junctions and Road Side Units (RSUs) are analysed by computer algorithms using either image processing or artificial intelligence-based techniques with respect to various traffic parameters like speed, density:

  • a.

    Image Processing Techniques: The image processing technique used for vehicle tracking and counting is background subtraction. It is a motion-based image segmentation method (Karandeep, 2017).

  • b.

    Deep learning techniques: After the advent of deep learning, accuracy for many computer vision tasks has increased by a large margin, especially for detection and classification. R-CNNs and You Only Look Once (YOLO) are some important state-of-the-art detectors. These are real time object detection systems which can detect the bounding box coordinates of all the vehicles in a given image. Trade-off between the speed and accuracy can also be achieved by changing the size of the model. The detected image patches can be further classified into various classes like cars, trucks, motorcycles, etc using deep neural network classifiers like VGG16, ResNet, DenseNet, etc (Evan, 2018).