HUMAN LOCALIZATION WITH MULTI-CAMERA USING DETECTION AND TRACKING OBJECT

• Main Author: WINARDI - NIM: 23515064 , SUNARYO
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/31141
• Institution: Institut Teknologi Bandung

<p align="justify">Objects Localization in the room, is a system that can determine the position of something or someone in the physical space. This system is widely used to perform human monitoring while in the room. Localization systems that have been designed using wireless sensor networks such as RFID, GPS, or Celluller networks are less effective in terms of cost because every object to be tracked must carry sensors that match the technology used. Therefore, systems that utilize the camera as low cost positioning existing ones will be developed. <br /> <br /> The system designed on this research is utilizing the input of multi-cameras to perform human localization. The inputs received by each camera will be processed and the results are combined to get the human location on the video. The result of combining the process is used to determine the human location of the video into a digital map depicted using homography. <br /> <br /> Determination of human location using camera requires human detection process on image. This process is used to determine the human position in the image if there are human objects in the image accurately and quickly. Challenges in human detection are environmental conditions (lighting and human density), camera capture (frame rate, non-human objects in input, and camera position) and other variations such as human poses, shapes, structures and movements. These challenges can result in undetected human detection and the occurrence of error detection of non-human objects detected as human beings. <br /> <br /> To solve the challenges, testing is done on the process of detection of human objects against multiple input cameras. The test is divided into three tests. First, human detection is done by searching the Histogram of Oriented (HOG) feature from input image. The HOG feature acquired will be classified using the Support Vector Machine (SVM). Results from SVM objects are classified as human in image. <br /> <br /> To reduce SVM errors in calcifying objects, preprocessing Multi-Scale Spectral Residual (MSR) is given to the image before object detection is performed. MSR is a process of subtraction search areas on images that can be used to speed up detection of features and reduce detection errors. <br /> <br /> Testing on this process is done by comparing SMV classification results using preprocessing MSR and without preprocessing. The goal is to get a camera that can provide the most effective <br /> <br /> human detection results. This most effective camera will be used in building a localized human system. <br /> <br /> The second test, comparing the speed of the image process. The speed of the image process in detecting and classifying features will be compared against images using MSR preprocessing and without preprocessing. In addition, the image processing speed will also be calculated from the input image received until the SVM classification is performed. Speed testing is performed to obtain the required average processing time of an image using preprocessing of MSR and without preprocessing in detecting human objects. <br /> <br /> Third, the system of localization of multi-camera human objects will be built by combining the most effective cameras of the first test. Combined results of the camera process will be tested. The test is done by calculating the accuracy of the mapping location on the digital map with the physical condition and the effectiveness level of object localization using multi-camera. <br /> <br /> The test results showed that the method without preprocessing MSR gives the efficiency result in detecting human average is 53.41% with the average process time is 0.16 seconds. While the use of preprocessing MSR provides average efficiency of human object detection 30.25% and average processing time 0.3-0.4 seconds. For the input camera, camere with True Positive = 210 and camera with True Positive = 216 provide the most correct detection results. With the above test, it can be developed multi-camera mapping system with HOG detection feature without preprocessing MSR using both cameras. <br /> <br /> Combining detection results from the simplified method of Error Bias Weighting use for determining the best detection result. By combining the results of both cameras, the efficiency of the camera mulit-mapping system was 87.07% with a minimum accuracy of 0.02 meters error and a maximum error of 2.2 meters from the human location point on the physical space.<p align="justify">