Video dizilerindeki araç plakalarının FE-yoğunlaştırma algoritması kullanılarak izlenmesi

Bu çalışmanın amacı, araç plakalarının üç boyutlu uzayda konum ve yöneliminin bulunması için video görüntüsünden izlenmesidir. Eğer nesnenin altı dereceli uzay serbestliği belirlenmek isteniyorsa, durum uzayı altı boyutlu olur. Her serbestlik derecesi için olası değerler kümesini 100 elemanlı kabul edersek, bu nesneyi her olası durumu deneyerek izleyebilmek için görüntü verisi üzerinde 1006 karşılaştırma yapmamız gerekmektedir. Bu sınırlı çözünürlük ve altı dereceli serbestlik uzayında dahi, bu şekilde gerçek zamanlı izleme yapmanın mümkün olmadığı açıktır. Stokastik izlemenin ardında yatan düşünce, her olası nesne durumunu denemek yerine, durum hakkında tahminlerde bulunmak ve bu tahminleri o anki video karesi ile karşılaştırarak sonuçları bir sonraki video karesi için tahmin yapmakta kullanmaktır. Son yıllarda, bilgisayar ile görüntü işleme problemlerinde Parçacık Filtreleri’nin kullanımına yönelik bir ilgi görülmektedir. Bilgisayar ile görüntü işleme problemlerinde kullanılan özel Parçacık Filtresi’ne Yoğunlaştırma algoritması veya Ardışıl Önem Örnekleme denmektedir. Bu yöntem hareketli nesneler için gürbüz bir izleme olanağı sunmaktadır. Öte yandan, bu algoritmanın yakınsaması büyük oranda parçacık sayısı ve dinamik modelin doğruluğu arasındaki ilişkiye bağlıdır. Bu tezde Yoğunlaştırma algoritmasını iyileştirmek amacıyla FE-Yoğunlaştırma algoritması önerilmektedir. Bu algoritma Farksal Evrim ve Yoğunlaştırma algoritmalarının bir birleşimidir. FE-Yoğunlaştırma algoritması üç boyutlu uzayda tek bir kamerayla araç plakası konum ve yöneliminin izlenmesi için kullanıldı. Genişletilmiş Kalman filtresi, Yoğunlaştırma, Genetik Yoğunlaştırma ve FE-Yoğunlaştırma algoritmalarının izleme başarımları karşılaştırıldı. FE-Yoğunlaştırma algoritması diğer üç algoritmaya göre çok daha iyi başarım göstermektedir.

Tracking of license plates in video sequences using DE-condensation algorithm

Automated vehicle identification (AVI) is still an important research issue and drawing attention in machine vision community. Its potential commercial applications are automatic barrier systems, automatic payment of parking or highway toll fee, automatic locating of a stolen vehicle, automatic calculation of traffic volume and so on. License plate enables us to identify a vehicle and its owner. License plate recognition is the most effective method for identification of the vehicle. A suitable and promising solution to vehicle identification is visual recognition of the license plate from camera view. This approach is applicable because it does not require vehicles to carry additional equipment such as special RF transmitters. Without additional cost, these systems are capable of installation to the field. But visual license plate detection and recognition is a very difficult task. It is quite a challenging problem because vehicles are running in an outdoor environment, where lighting conditions can change rapidly, weather conditions can cause poor image quality, license plates can be dirty or in poor condition and occlusions can occur frequently. Therefore, Visual License Plate Recognition (VLPR) systems may fail because of uncontrollable external conditions. Beside the challenging nature of the problem, the high-dimensional nature of the VLPR problem may impose a significant computational load on the target processing platform. The aim of this work is 3D tracking of license plate in order to determine the state (spatial position and 3D orientation) of the license plate from sequential frames of the video. This can be accomplished in a brute force way by testing every possible orientation and translation and then selecting the one that best fits the current frame. If all six degrees of spatial freedom of the object are to be determined, the state space of the object is six dimensional. Setting the number of possible values of each degree of freedom to 100, the task of tracking by brute force then requires 1006 comparisons of a state with the image data. Even with such a limited resolution and a six dimensional feature space it is clear that, it is computationally impossible to perform tracking in real time by brute force. That is where stochastic tracking is meaningful. A stochastic process is one whose behaviour is non-deterministic in that the next state of the environment is partially but not fully determined by the previous state of the environment. Instead of comparing every possible configuration of the object with each video frame, the idea behind stochastic tracking is to make a set of guesses of the state, compare these guesses with the current frame, and use the result of this comparison as the basis for a new set of guesses when the next frame comes. The new guesses are made by selecting the best guesses from the last frame and applying a model of the movement of the object from one frame to the next. The set of guesses (called particles or samples) will frame by frame converge around the correct state of the object. In recent years, there has been a great interest in applying Particle Filtering to computer vision problems. This specialized Particle Filtering method for computer vision problems is introduced as Condensation or Sequential Importance Sampling. Condensation algorithm utilizes factored sampling and given dynamic models to propagate an entire probability distribution for object’s position and shape over time. It can perform successfully robust tracking of object motion. On the other hand, its convergence greatly depends on the trade off between the number of particles/hypotheses and the fitness of the dynamic model. For example, in cases where the dynamics are complex or poorly modelled, thousands of samples are usually required for real applications. In order to improve the performance of the Condensation algorithm, DE-Condensation algorithm is proposed, which is an integration of the Differential Evolution and Condensation algorithms. DECondensation algorithm is utilized for spatial position estimation and tracking of license plates in 3D from monocular camera view. The performance and computational load of the Extended Kalman filter, Condensation Algorithm, DE-Condensation algorithm and Genetic Condensation algorithm are compared for evaluating DE-Condensation Algorithm’s performance.

___

  • Anagnostopoulos, C., Anagnostopoulos, I., Tsekouras, G., Kouzas, G., Loumos, V., Kayafas, E., (2005). Using sliding concentric windows for license plate segmentation and processing, Proceedings, Signal Processing Systems Design and Implementation, 337–342, Athens.
  • Ching-Tang, H., Yu-Shan, J., Kuo-Ming, H., (2005). Multiple license plate detection for complex background, Proceedings, 19th International Conference, Advanced Information Networking and Applications AINA, 2, 389 – 392, Taipei.
  • Dubey, P., (2005). Heuristic approach for license plate detection, Proceedings. IEEE Conference on Advanced Video and Signal Based Surveillance, 366 – 370, Como.
  • Feoktistov, V., (2006). Differential Evolution In Search Of Solutions, Springer, Berlin.
  • Hsi-Jian, L., Si-Yuan, C., Shen-Zheng, W., (2004). Extraction and recognition of license plates of motorcycles and vehicles on highways, Proceedings, 17th International Conference, Pattern Recognition, ICPR 2004, 4, 356 – 359, Cambridge.
  • Isard, M. ve Blake, A., (1998). A mixed-state condensation tracker with automatic modelswitching, Proceedings, Sixth International Conference on Computer Vision, Jan., 107 – 112, Bombay.
  • Price, K. V., Storn, R. M. ve Lampinen, J. A., (2005). Differential Evolution a Practical Approach to Global Optimization, Springer, Berlin.
  • Ponsa, D., Lopez, A., Serrat, J., Lumbreras, F., Graf, T., (2005). Multiple vehicle 3D tracking using an unscented Kalman Filter, Proceedings, Intelligent Transportation Systems, 1108 – 1113, Vienna.
  • Sang Kyoon, K., Dae Wook, K., Hang Joon, K., (1996). A recognition of vehicle license plate using a genetic algorithm based segmentation, Proceedings , International Conference on Image Processing, 1, 661 – 664, Lausanne.
  • Vacchetti, L., Lepetit, V., Fua, P., (2004). Combining edge and texture information for real-time accurate 3D camera tracking, Proceedings, Third IEEE and ACM International Symposium Mixed and Augmented Reality, 48 – 56, Arlington.
  • Wen-Yan, C., Chu-Song, C., Yi-Ping, H., (2005). Appearance-guided particle filtering for articulated hand tracking, Proceedings, IEEE Computer Society Conference Computer Vision and Pattern Recognition, 1, 235 – 242, San Diego.
  • Wenjing, J., Huaifeng, Z., Xiangjian, H., Piccardi, M., (2005). Mean shift for accurate license plate localization, Proceedings, Intelligent Transportation Systems Proceedings, 566 – 571, Vienna.
  • Xiong, J., Du, S., Gao, D., Shen, Q., (2004). Locating car license plate under various illumination conditions using genetic algorithm, Proceedings, 7th International Conference on Signal Processing ICSP '04, Volume 3, 2502 – 2505, Beijing.
  • Zayed, M., Boonaert, J., Bayart, M., (2004). License plate tracking for car following with a single camera, Proceedings, The 7th International IEEE Conference on Intelligent Transportation Systems, 719 – 724, Washington.
  • Zhu, Y., Zhi-Qiang, L., (2005). Genetic CONDENSATION for motion tracking, Proceedings, International Conference on Machine Learning and Cybernetics, 9, 5542 – 5547, Guangzhou.