Stereo Matching algorithm

First, the classification of stereo matching algorithm

In stereo matching, the matching problem can be regarded as the process of finding the correlation degree of two sets of data. Stereo matching algorithm is classified by many kinds.

① is based on the scope of the algorithm runtime constraint: The local matching algorithm and global matching algorithm.

② based on the generated parallax map: It can be divided into dense (dense) matching and sparse (Sparse) matching. Dense matching: is based on the generated parallax map, which can generate a deterministic parallax value for all pixels, called dense matching. Sparse matching: Select only the key pixels [usually corner or edge points] The method of calculating parallax value is called sparse matching, the algorithm is fast, but it also needs to calculate the parallax value of missing pixels by interpolation algorithm, so there are a lot of limitations in the application scenario.

As a result of their recent research focused mainly on local matching algorithms and global matching algorithms, the following will also be described here.

1. Local matching algorithm and global matching algorithm

Global (semi-global) stereo matching algorithm mainly uses the global optimization theory method to estimate parallax, establishes a global energy function, which contains a data item and a smoothing term, and obtains the optimal parallax value by minimizing the global energy function. Among them, graph cutting (graph cuts, GC), confidence Propagation (belief PROPAGATION,BP), dynamic Programming (dynamical PROGRAMMING,DP), particle swarm optimization (particle Swarm optimization, PSO), Genetic algorithm (genetic algorithm,ga) and other optimization algorithms are commonly used to solve the energy minimization method.

The local stereo matching is the same as the global stereo matching algorithm, and the optimal parallax is calculated by optimizing a cost function method. However, in the energy function of local stereo matching algorithm, only constrained data items based on local region are not smoothed. The local matching algorithm only uses the gray, color, gradient and other information of a certain neighborhood to calculate the matching cost, the computational complexity is low, most real-time stereo matching algorithm belongs to the category of local stereo matching, but the local stereo matching algorithm is not ideal for low texture region, repeated texture region, parallax discontinuity and occlusion region matching effect.

Two, stereo matching algorithm evaluation platform

①middlebury Test Platform: Provides a test image pair (Stereo pair) specifically for evaluating stereo matching algorithms, including Tsukuba test image pair, Venus test image pair, Teddy test image pair and cones test image pair, with resolutions of 384 *288,434*383 and 450*375, the real disparity maps of these test images are also given.

②kitti Algorithm Evaluation platform: Designed to evaluate the object (motor vehicles, non-motor vehicles, pedestrians, etc.) detection, target tracking and other computer vision technology in the vehicle environment performance, for motor vehicle-assisted driving should be used as technical assessment and technical reserves. Kitti contains real-world image data collected from scenes such as urban, rural and expressway

Stereo Matching algorithm