Real-Time Stitching Algorithm of Vehicle Side View Image Based on Multi-Region Fast Phase Correlation

Abstract

Aiming at the problem that the camera field of view is too small to obtain the complete side image of the vehicle in the traffic management scene, a real-time stitching algorithm of vehicle side view image based on Multi-Region Fast Phase Correlation (MFPC) is proposed. Firstly, the background subtraction method based on Gaussian mixture model is used to obtain vehicle foreground image sequence, and the image is downsampled by Gaussian pyramid to reduce the running time of the program. Subsequently, multi-region phase correlation and registration check based on normalized cross-correlation are used to improve the accuracy of registration, and a local inverse discrete Fourier transform method is proposed to improve the computational efficiency. To mitigate background interference in registration, a peak filtering algorithm is proposed, combined with a sub-pixel refinement algorithm to enhance accuracy. Experimental results indicate that the proposed algorithm has better registration performance than the traditional phase correlation and methods based on feature point detection. It requires only 26% of the time taken by the traditional phase correlation, with an average processing time per frame of 3.81 ms, which meets the real-time requirements. In practical applications, the stitching accuracy reached 99.33%, demonstrating high precision and robustness.

Keywords

• Image registration
• image stitching
• phase correlation
• road vehicles
• vehicle detection