A High-Performance Moving Object Detection Method Based on Optical Flow

Abstract

An adaptive and high precision optical flow estimation approach for moving object detection is proposed. The proposed method (P-M) is composed of a K-means clustering based particle swarm optimization algorithm (PSO-K), an improved multi-scale method and a flow field verification strategy. To test the P-M, a series of experiments are carried out. The experimental result based on the Middlebury training set shows that the P-M estimates the uniform distribution of flow field and the boundary between moving objects is clearly visible. Moreover, the P-M has the highest accuracy with minimal average endpoint error (AEPE) and average angular error (AAE) compared to the Lukas Kanade (LK) method, the classic Horn Schunck (C-HS) method and block-based matching (BL) method. The AEPE and AAE for the P-M are 0.427 and 3.402, respectively. The maximum average relative improvement rates (ARIR) are 43.816% and 70.252 %, respectively. Furthermore, the test result of the micro-vision image sequence demonstrates that the P-M has a high performance, which can accurately detect the moving targets even in the presence of large displacement.