EVELOPMENT OF MEAN-SHIFT BASED FACE TRACKING SYSTEM WITH LOCUST SEARCH ALGORITHM OPTIMIZATION

• Main Author: Retno Hardini, Inkreswari
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/53308
• Institution: Institut Teknologi Bandung

The purpose of object tracking is to identify objects, determine position, and update the object's position (region of interest) continuously. Object tracking systems were applied to various things, one example of application is in the localization of postdisaster victims. Based on this, having an optimal object tracking system is a necessity. One form of the optimal object tracking system is a system that can achieve the ROI convergence point in each video frame with a minimal number of iterations so that it has an impact on the system process in following the movement of the target object to be fast. The number of iterations generated in reaching the convergent point of ROI is directly proportional to the system's ability to follow the object's movement. The greater the number of iterations generated, the slower the system will follow the object's movement. Conversely, the smaller the number of iterations, the faster the system will follow the movement of objects. Therefore, having a system that can follow the movement of objects quickly, concerning its application in the field of post-disaster victim localization, is urgently needed. To track objects in this study, the Mean-Shift algorithm is used. The Mean-Shift algorithm technique in finding the convergent point of ROI in each frame will shift the position of the point sequentially until reaching the convergent point of ROI. This has an impact on the number of iterations required is large and has an impact on the system process in following the movement of objects to be slower. To overcome this, the optimization algorithm is used with the aim of finding the convergent point of ROI faster so that it has an impact on the number of iterations required is less and has an impact on the system process in following object movements to be faster. The Simulated Annealing optimization algorithm in its application to find the convergent point of ROI still has shortcomings from its search technique. The technique that is owned is that at each iteration a new random point will be built and the function value of that point is calculated, so it is likely that the optimum point obtained at this time is a local optimum point. This happens because Simulated Annealing does not have the ability to find a position or build a new random point based on the previous optimum point position. If the built-in points are random, finding the global optimum point will require more iterations which will result in longer processing time. This study aims to produce an object tracking system with an object tracking algorithm that has a faster processing time kinerjance compared to the Mean-Shift algorithm that is optimized with Simulated Annealing. The proposed treatment is to replace the Simulated Annealing optimization algorithm with the Locust Search algorithm. The Locust Search algorithm was chosen because of its technique, which is that each iteration will build a new point to find the maximum or optimum point by paying attention to the position of the previous optimum point. This has the advantage that the number of iterations required is much less and has an impact on the faster process of following the movement of the target object. The research was implemented on the Raspberry Pi 3 Model B + in the form of a prototype. The results of the kinerjance evaluation use 500 image frame data from the video file with experimental iterations for each tracking algorithm is 20 experiment iterations. So that, the total number of experiments for the all of 3 algorithms are 60 iterations, uses the one-tail t-test testing technique with the assumption of different variants. The results obtained show that the proposed algorithm has a better average kinerjance than the Mean-Shift algorithm and the Mean-Shift algorithm that is optimized with Simulated Annealing, which is 151 milliseconds.