, pp.32-36 http://dx.doi.org/10.14257/astl.2016.129.07 Multi-Resolution Estimation of Optical Flow on Vehicle Tracking under Unpredictable Environments Viet Dung Do 1 and Dong-Min Woo 1 1 Department of Electronics Engineering, Myongji University, Yongin, Republic of Korea mr.dovietdung@gmail.com and dmwoo2000@gmail.com Abstract. This paper proposes a vehicle tracking system that can effectively solve a multi-resolution problem under unpredictable environments. Our system uses Scale-Invariant Feature Transform (SIFT) algorithm as the feature detector. The pyramidal Lucas Kanade optical flow algorithm using our feature tracking system is then implemented. The information after the estimation of optical flow can be used in the later tracking and detecting problems. For the purpose of evaluation, we choose the Autonomous Agents for On-Scene Networked Incident Management (ATON) project s highway video files which include moving shadows. Experimental results show good performance of our multi-resolution optical flow system. It is confirmed that the computed optical flow has very small errors in minimum eigenvalues of optical flow equations. Keywords: vehicle tracking, optical flow, multi-resolution, Scale-Invariant Feature Transform, feature tracking 1 Introduction Vehicle tracking is the process of locating a moving object or multiple objects using a camera. There is many image processing techniques for this object tracking problem [1]. Some real difficulties of vehicle tracking are tracking vehicles moving with extremely high speed, tracking a huge number of vehicles, or tracking under unpredictable situations. These situations can suddenly appear at any time, from data capturing, data processing or data storing. Data which plays an important role in this field should be in images or video files. Good quality data helps scientists implement algorithms, experiment on hypotheses, build novel processes or systems, estimate the accuracy and make comparisons with other systems. Because of this important role, data selection should be concerned carefully [2]. One of real difficulties in vehicle tracking is a multi-resolution problem. A multiresolution problem appears to be becoming more common when working with data in multi-scale and multi-view. Another popular example of multi-resolution problem is noise reduction. This is the reason why we need to pre-process the data in order to add noise in input video files. Experiments in [3][4] show promising results of correlationbased optical flow under noisy environments. In this paper, experiments are examined ISSN: 2287-1233 ASTL Copyright 2016 SERSC
not only in noisy environments but also in vibration and blurring environments. The results show that our application has good accuracy with fast computation time. In order to reduce the computation time, optical flow algorithms should only be computed on interest points. Interest points [5] which are a well-defined position can be detected with affine transformations including translation, rotation, scaling etc. One of efficient corner detectors is Scale-Invariant Feature Transform (SIFT) [6]. Features detected by SIFT is useful in solving the problem between multi-views of an object or a scene, and invariant to affine transformations, even addition of noise and motion blur. After detecting strong interest points, we apply the Shi and Tomasi method [6] to track good features. By our experiments, this method will been proven effective in combination with optical flow algorithms in motion tracking generally and vehicle tracking privately. 2 The Tracking System Fig. 1 shows our overall tracking system. Our application takes video files or sequences of images as an input data. These inputs should be recorded the highway traffic by a fixed camera. The video files that are used in our experiments have moving shadows from vehicles. This will challenge our testing more. Fig. 1. Overall vehicle tracking system The pre-processing step consists of vibration effect generation, noise generation and blurring. The main goal of our application is testing tracking system under various environments. So we need to add more disadvantage effects in our inputs in order to test the system. The core of our application includes two parts: feature tracking and optical flow estimation. We apply the Scale-Invariant Feature Transform (SIFT) algorithm as a corner detector in order to detect strong interest points. After that, we can use these points as parameters in Shi and Tomasi feature tracking system. The optical flow estimation step takes the results from previous step and computes the optical flow. Copyright 2016 SERSC 33
We implement the iterative Lucas Kanade algorithm in combination with pyramid scheme in order to solve the multi-resolution problem. 3 Multi-Resolution of Optical Flow Image pyramid is widely used as architecture to solve the multi-resolution problem. An image pyramid is a collection of images all arising from a single original image that are successively sampled until some desired stopping point is reached. There are two types of image pyramid: the Gaussian and the Laplacian pyramids [7]. The Gaussian pyramid is used to down-sample images, while the Laplacian pyramid is used to reconstruct an up-sample image from an image lower in the pyramid. Fig. 2 illustrates three level pyramid of one frame of the video input file used in the experiment. Fig. 2. Three level pyramid. One of the difficulties when implementing Lucas Kanade algorithm is that points might be moved outside of the local windows. Then it is hard for the algorithm to find these points. This leads to the pyramidal Lucas Kanade algorithm which works starting from the highest level of an image pyramid and keep going down to lower levels. When combining the pyramidal Lucas Kanade optical flow algorithm with image pyramid after feature tracking system based on SIFT, the performance will be improved in both accuracy and time cost, especially in a multi-resolution problem. 4 Experiment We examine highway camera video files from Autonomous Agents for On-Scene Networked Incident Management (ATON) project. Our application also detects the moving shadow of vehicle in these two video files, leaves the moving shadow, and keep the motion of vehicles. The numbers of frames of the file are 440 and 500. Each frame of video file is artificially corrupted by two types of noise, additive Gaussian noise and additive Laplacian noise. The results are shown in Fig. 3. 34 Copyright 2016 SERSC
The results show that our tracking system can detect all of moving vehicles. However, there are small features from moving shadows which cannot be eliminated. We examined these experiments on Intel Core i5-3570, 3.40 GHz under Microsoft Visual Studio 2010 C++ environment. Table 1 shows the computation time on two highway video files. (a) (b) (c) Fig. 3. Vehicle tracking with (a) Additive Gaussian noise at frame 12 of 440. (b) Additive Gaussian noise at frame 179 of 440. (c) Additive Laplacian noise at frame 139 of 440. (d) Additive Laplacian noise at frame 356 of 440. Table 1. Computation time on vehicle tracking under unpredictable environments. Input Environment highwayi_raw.avi (440 frames) highwayii_raw.avi (500 frames) Additive Gaussian Noise 5.891 7.338 Additive Laplacian Noise 6.048 8.567 The experimental results show that our application has good performance with small errors when estimating the optical flows, even under unpredictable environments. The accuracy of highwayii_raw video files is better, because the moving shadows in this input video are smaller than in highwayi_raw files. (d) 5 Conclusion We proposed a moving tracking system of moving vehicle based on optical flow and feature tracking in a multi-resolution fashion. The results show good performance when working with input data even in unfavorable environments with noise. Although Copyright 2016 SERSC 35
most of moving shadow is eliminated, some small features of shadow are still undetected. As an extension of this work, we need to solve more complex problems in motion tracking including vibration, blurring and lighting changes, and evaluate the performance, along with the proposed tracking system in this paper. Acknowledgments. This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. NRF-2012R1A1A2004950). References 1. Ojha, S., Sakhare, S.: Image Processing Techniques for Object Tracking in Video Surveillance A Survey. International Conference on Pervasive Computing, pp.1--6 (2015) 2. Li, S., Yu, H., Zhang, J., Yang, K., Bin, R.: Video-Based Traffic Data Collection System for Multiple Vehicle Types, IET Intelligent Transport Systems, 8, 164--174 (2013) 3. Kesrarat, D., Patanavijit, V.: Experimental Efficiency Analysis in Robust Models of Spatial Correlation Optical Flow Methods under Non Gaussian Noisy Contamination, International Conference on Electronics, Computer, Telecommunications and Information Technology (2013) 4. Lee, T., Anderson, D.: Performance Analysis of a Correlation-Based Optical Flow Algorithm under Noisy Environments, IEEE International Symposium on Circuits and Systems, pp. 4699--4702 (2006) 5. Shi, J., Tomasi, C.: Good Features to Track, IEEE Conference on Computer Vision and Pattern Recognition (1994) 6. Lowe, D.: Distinctive Image Features from Scale-Invariant Keypoints, International Journal of Computer Vision, 60, 91--110 (2004) 7. Burt, P., Adelson, E.: The Laplacian Pyramid as a Compact Image Code, IEEE Transaction on Communication, 31, 532--540 (1983) 36 Copyright 2016 SERSC