A Video Image Processing Method for Continuous Object Detection

Wenting Ma, Xiaohang Yuan, Jian Zhu, Haojie Zhang, Zhipeng Zhang, Yuan Hao

Abstract: Post-processing of video images is essential to the whole video image detection, especially for continuous objects. Continuous objects refer to objects with continuity, integrity, and consistency at the level of physical media or data application, including tracks, cables, lane lines, chain structures, etc. Usually, objects are composed of a series of small homogeneous units, such as a piece of track, a piece of cable, a color ring, etc. The rapid development of artificial intelligence technology and 5G communication technology has driven the gradual maturity of deep learning and autonomous driving technology. At the same time, the standard for the detection of continuous objects is growing higher and higher. Therefore, how to detect objects in different business scenarios and scene-based optimization have become a top priority. The existing problems include: (1) Small units of multiple continuous objects in complex scenes interfere with each other, hindering recognition; (2) Various negative factors (insufficient brightness, steam, smoke occlusion, etc.) cause poor image quality. Consequently, the recognition performance is negatively influenced, and the model cannot adjust parameters adaptively; (3) The skew and distortion of image lower the object recognition performance. Aiming at these problems, this paper proposes a video image post-processing optimization algorithm for object detection of continuous objects, including an object integration search algorithm, threshold adaptive adjustment algorithm, rotation perspective correction algorithm, etc. Prior knowledge is utilized to improve the accuracy and efficiency of detection.