Matching techniques to compute image motion
Introduction
Window-matching or correlation-based techniques are the most intuitive and perhaps also the most widely applied techniques to compute the optical flow from an image sequence, i.e. to estimate the 2D motion projected on the image plane by the objects moving in the 3D scene [1], [2], [3], [4], [5], [6]. Optical flow estimation has many practical and industrial applications, i.e. for object tracking, assisted driving or surveillance systems, obstacle detection, image stabilisation or video compression [7], [8], [9], [10], [11]. In spite of this fact, few works analysing the performances and the possible enhancements of these algorithms have been presented [1], [2], [6] so that a more detailed analysis of this simple and widely used optical flow technique seemed to us necessary. The aim of this paper is to give a clear overview of window matching algorithms and to suggest new solutions to improve on their shortcomings (such as the computational cost, the pixel precision, and so on).
The paper is organised as follows: Section 2 gives an overview of correlation-based techniques discussing advantages and drawbacks, Section 3 introduces the distance or similarity measures we applied to our algorithms. Section 4 discusses the matching error due to high frequencies and search space quantisation. Section 5 introduces several techniques in order to have the best results in matching, reducing the complexity and increasing the accuracy obtaining also a sub-pixel motion estimation. Section 6 presents the experimental results, with comparisons of algorithms on well-known test image sequences.
Section snippets
Overview: advantages and drawbacks
Correlation-based methods are based on the analysis of the gray level pattern around the point of interest and on the search for the most similar pattern in the successive image. In a few words, having defined a window around the point we consider similar windows W′(x+i,y+j) shifted by the possible integer values in pixels in a search space S composed by the i, j such as −Δ<i<Δ and −Δ<j<Δ. The optical flow, i.e. the estimated image displacement is taken as the shift corresponding to
Distance–similarity measures
Many ways of measuring difference or similarity between the gray-level pattern can be used. In our work we compare squared windows of N×N and compute motion between a window centred in (x,y) in the image I1 and a window shifted by (i,j) in the image I2. The most used distance measures are reported in Table 1. The widely used sum of absolute differences (SAD) and SSD can be modified to consider the effect of global gray-level variations, setting the average gray level difference equal to 0
High frequencies and quantisation
Even if correlation-based techniques are not affected by the aliasing problem as differential ones, signal quantisation introduces error in flow computation due to high frequencies. If the frequency of the signal has the same order of magnitude as the sampling frequency and the displacements to be computed are not exactly integer (i.e. a multiple of the sampling step), correlation may lead to completely wrong results as well. Let us show it with a simple example: we consider a 1D sinusoidal
Improving the method
The previous sections have shown several features of correlation algorithms also showing some problems of the method; in this section we will introduce several algorithms to solve these problems and to improve the algorithm performances.
Experimental results
To test the algorithms, we computed optical flows on synthetic or calibrated image sequences with the true displacements known at every pixel location. We measured the average differences between the computed flow and the true motion using the angular distance introduced by Barron et al. [2]:
Discussion
The use of techniques based on pattern matching performed on image regions at different times is common in practical applications, but a few algorithms of this kind are considered in the literature reviews [2]. In this paper several variations of this kind of algorithms and some tricks to reduce the major drawbacks of the method (i.e. the computational complexity and the integer values of the estimates) are presented. All the solutions proposed have been tested on the test images used by the
Acknowledgements
Special thanks to Pietro Parodi for reading the manuscript and useful hints and to Marco Cappello and Marco Campani for technial help. Thanks also to Vincent Torre, Alessandro Verri for useful discussions.
References (19)
- et al.
Determining optical flow
Artificial Intelligence
(1981) Displacement vectors derived from 2nd order intensity variations in image sequences
Comput. Vision Graph. Image Process
(1983)Obstacle detection by evaluation of optical flow fields from image sequences
IVC
(1991)On line analysis of echocardiographic images sequences
Med. Image Anal.
(1998)computational framework and an algorithm for the measurement of visual motion
Int. J. Comput. Vision
(1989)- et al.
Performance of optical flow techniques Int
Int. J. Comput. Vision
(1994) - et al.
An interative image registration technique with an application to stereo vision
Proc. DARPA Image Und. Workshop
(1981) - R. Singh, An estimation-theoric framework for image flow computation, Proc. Third ICCV, Osaka, 1990, pp....
- et al.
Optical Flow from 1D Correlation: application to a simple time-to-crash detector
Int. J. Comp. Vision
(1995)
Cited by (194)
Early-age shrinkage assessment of cementitious materials: A critical review
2024, Cement and Concrete CompositesMeasuring landslide-driven ground displacements with high-resolution surface models and optical flow
2023, Computers and GeosciencesInteractive Structural Topology Optimization with Subjective Scoring and Drawing Systems
2023, CAD Computer Aided DesignMulti-path decoder U-Net: A weakly trained real-time segmentation network for object detection and localization in ultrasound scans
2023, Computerized Medical Imaging and GraphicsDIC-Net: Upgrade the performance of traditional DIC with Hermite dataset and convolution neural network
2023, Optics and Lasers in EngineeringA state-of-the-art review of mechanical characteristics and cracking processes of pre-cracked rocks under quasi-static compression
2022, Journal of Rock Mechanics and Geotechnical EngineeringCitation Excerpt :These subsets contain many pixels which can be used for distinguishing themselves from others (Pan et al., 2009). To determine the correlation degree of subsets between reference and deformed images, the correlation criteria were developed (Grün, 1985; Giachetti, 2000; Pan et al., 2010). When the position with the maximum correlation coefficient was determined, the relationship between the reference and deformed images was identified, and the full-field displacement and strain could be subsequently calculated (Blaber et al., 2015).