Real time classification and tracking of multiple vehicles in highways

Abstract

Real time road traffic monitoring is one of the challenging problems in machine vision, especially when one is using commercially available PCs as the main processor. In this paper, we describe a real-time method for extracting a few traffic parameters in highways such as, lane change detection, vehicle classification and vehicle counting. In addition, we will explain a real time method for multiple vehicles tracking that has the capability of occlusion detection. Our tracing algorithm uses Kalman filter and background differencing techniques. We used morphological operations for vehicle contour extraction and its recognition. Our algorithm has three phases, detection of pixels on moving objects, detection of a “Shape of Interest” in frame sequences and finally determination of relation among objects also in frame sequences. Our system is implemented on a PC with Pentium II 800 MHZ CPU. Its processing speed was measured to be 11 frames per second. The accuracy of measurement was 96%.

Introduction

High speed processing of image frame sequences is highly important for many real time computer vision algorithms. Image sequence analysis provides intermediate results for conceptual description of events in a scene. In vehicle tracking applications that uses image sequences, many methods are suggested. At first we can name model-based tracking methods in which a 3D model of vehicle is extracted (Coifman et al., 1998, Malik and Russell, 1997). One of the advantages of these methods is their high accuracy in determining the vehicle type and their detail geometric model. In fact, model-based tracking methods because of their high calculation cost, can be used only for free-flowing traffic with small number of vehicles.

In some other methods that are called feature-based (Roberts, 1994), a few features such as distinguishable lines or corners are extracted for each vehicle. Some of these features are grouped together to label a vehicle (Shi and Tomasi, 1994). One of the most important advantages of this type of methods is that, even in presence of partial occlusion, some of these features remain to be visible. On the other hand, they face problems in detecting features for individual vehicles that run close to each other. In region-based methods (Coifman et al., 1998, Setchell, 1997), vehicles are presented as blobs. In these methods, at first, connected components are extracted and then regions are merged or split ted if needed. The most serious weakness of these approaches is that merging and splitting regions can cause some inaccuracy in vehicle detection.

In addition, there are some methods in which the contour of vehicles is extracted. Although contours can be detected by simple edge detection methods, but these simple methods sometimes detect false edges of the background too. However, if more complex algorithms of edge detection such as active contours or snakes (Paragios and Deriche, 2002) are used, one should find a way to optimize the coding to make them usable for real time applications with commercially available processors. In practice there are many applications such as our system, that often one does not need to know the exact detail of vehicle type, but a general type category would be sufficient. In our system the surveillance CCD camera is installed in a relatively far distance from a highway and the vehicles are visualized as small objects with minimum detail on their geometrical model.

In this paper we introduce a novel real time machine vision system for classification and tracking of multiple vehicles and also determining some traffic parameters such as lane change and counting the number of vehicles passing the highway during a desired time interval. For tracking we used Kalman filter (Grewal and Angus, 1993) and background differencing techniques. Our algorithm takes advantage of region based and contour based methods by combining their ideas in order to detect a “Shape of Interest”, that in practice it is a bounding box around the vehicle. By using the bounding box and region boundary, the occlusion and overlapping of two regions are detected by examining the object shape and determining if it was the result of merging two or more vehicles and then deciding upon a proper split point to separate the merged vehicles.

Our system was implemented in Visual C++ using Matrox Meteor II frame grabber on a Pentium II 800 MHZ CPU. The input images were gray scale with eight bits per pixel resolution and of size 320 × 320. Our experimental results showed an accuracy of 96% when it was compared with the measurements done by a human expert. The initial version of this work is given in (Rad and Jamzad, 2003).

The rest of this paper is organized as follows: After a review of related works, our algorithm is described in three main sections, Change detection, Vehicle recognition (where our ideas for occlusion removal and vehicle classification are discussed) and Vehicle tracking. In Section 6, the experimental results is presented and finally the conclusion remarks is given in Section 7.