Identification of Small Scale Objects During Large Scale Public Events - Assessment and Interaction of Image Segmentation and Machine Learning

Abstract

Public events like “Rock at the Ring”, “Wacken Open Air” or “Refugee Camps” and disasters are challenging situations for organizations with security responsibilities. In emergency situations, an effective and coordinated response for civil security is the availability of situational information and rescue forces. This can be considered as a requirement to ensure a quick and coordinated response that implies a fast analysis. This thesis focuses on Object-Based Image Analysis (OBIA) processes starting with image segmentation, the detection of small scale objects, followed by an object-based supervised image classification. The segmentation process is realized with three different methods, Multi-Resolution Segmentation, Mean-Shift Segmentation and Watershed Transformation. Image segmentation quality is carried out and evaluated by means of different characteristics like robustness and simplicity as well as performance and parameter transfer between different data sets and resolutions. The objective is to maximize homogeneity within the segments that to date ensure good classification results with least manual user interaction. Image segments are classified using Support Vector Machine and Random Forest Algorithms. Furthermore, classification accuracy is assessed using independent reference data. The relation and bridge between segmentation and classification is rated through the calculation of segments which are used as base for the classification. Spectral, spatial and textural features are considered for different orthophotos and satellite images. Challenging is the over- and under-segmentation of images which joins segmentation and classification results. Compared to other approaches, OBIA methods illustrate that image resolution indicates dependencies to segmentation quality which again leads to accurate classification results. Over-segmented images point to good classification accuracies although quality metric values are indifferent. The values of segmentation quality range between 0.9 (inadequate) and 0.19 (nearly ideal). The overall segmentbased classification accuracy has an average of 88% using the Support Vector Machine method and 93% using the Random Forest algorithm. A slightly over-segmented image separates neighboring objects in a natural way and recognizes all objects of interest without image under-segmentation. An object count after the classification process of orthophotos shows an overestimation and when using satellite images it turns into an underestimation. As perspective of this thesis, post-processing is needed for the analysis and interpretation of the object amount by a rule-based merging process.