Vehicle Recognition Based on Object Based Analysis of Airborne Remote Sensing Images

Vehicle detection and counting is an important issue for many applications such as remote monitoring, vehicle tracking for security purposes, traffic management, rescue tasks, parking capacity analysis and metropolitan planning. Vehicle monitoring is also an important part of traffic information, crash control, vehicle flow statistics, road network planning and parking position estimation. Remote sensing images are widely used to monitor vehicles, due to the ability of sensors in providing a complete coverage of the area of interest. Compared with satellite imagery, aerial imagery is usually more considered for vehicle detection and traffic monitoring purposes due to the higher spatial resolution. However, this is extremely challenging due to the small size of vehicles, their different types and orientations, and the visual similarity to some other objects, such as air conditioning in buildings, trash cans and road signs in high resolution images. Lots of researches have been carried out on vehicle recognition in aerial images over the past years. These works can be categorized into the two main groups; shallow learning based methods and deep learning based methods. Most of the researches proposed in the deep learning category use Convolution Neural Network (CNN) for automatic features extraction. Although local convolution neural networks have performed well in object recognition from images, their performance in aerial imagery is limited due to the small sizes and orientation of vehicles, the complex background in urban areas, and difficulties in rapid detection due to large covering area. The general strategy that is applied in shallow learning based methods relies on hand crafted features extraction followed by a classifier or cascade of classifiers.

In this paper, a shallow learning based vehicle recognition algorithm is proposed for aerial imagery. This method uses the advantages of object based image analysis and the image pyramid. The proposed automatic vehicle recognition algorithm is a decision fusion strategy between the initial vehicle candidates and land use/cover classification map to modify vehicle recognition results. The initial vehicle candidates are recognized by structural object classification based on image pyramid. The proposed algorithm for initial vehicle candidates generation is composed of four main steps; 1) generating image pyramid, 2) performing image segmentation on the pyramid layer, 3) structural features measurement on the segmented image objects of pyramid layer and 4) performing knowledge based classification of the image segments into the vehicle and no-vehicle classes to produce a binary map containing only the initial candidates of vehicles. The land use/cover classification map is also generated in an object based image analysis procedure. In the final step of the proposed automatic vehicle recognition in this paper, a decision fusion algorithm is performed between initial vehicle candidates and the generated land use/cover classification map. In this procedure, the recognized initial vehicle candidates from pyramid layer should be transferred to the original image resolution by performing inverse pyramid transformation. Then, considering the meaningful neighboring relationships between vehicles and other defined object classes, the final and modified vehicle regions are recognized.

The ability of the proposed vehicle recognition method in this study is evaluated based on Ultracam aerial imagery with spatial resolution of 11 cm and four spectral bands in visible and NIR that is taken from an urban area in southwestern Russia. The extent of this study area is about 5900 to 9100 pixels. The obtained results showed the vehicle recognition accuracy for about 80%. Moreover, %78.87 and 0.71 are respectively the values for overall accuracy and Kappa coefficient of the final classification map from the proposed decision fusion algorithm. The decision fusion algorithm can decrease false positive pixels in the vehicle recognition results by performing reasoning rules based on the relationships between vehicles and other objects such as buildings and roads.