Detecting smoky vehicles from traffic surveillance videos based on dynamic features

摘要

Existing smoky vehicle detection methods are vulnerable to false alarms because of the continuous interferences from common passed vehicles and the complex characteristics of smoke. This paper presents a video smoky vehicle detection method based on dynamic features. Three groups of features, including Multi-Sequence Integral Projection (MS-IP), Center-Symmetric Local Binary Patterns on Three Orthogonal Planes (CSLBP-TOP) and Histograms of Oriented Optical Flow (HOOF), are proposed or employed to characterize dynamic features of successive Region of Interest (ROIs). More specifically, the MS-IP characterizes the diffusion and distribution information based on multiple-sequence analysis and integral projection. The CSLBP-TOP characterizes the spatiotemporal texture information by (1) combining the strengths of Shift-Invariant Feature Transform (SIFT) and LBP and (2) extending the spatial features to three-dimensional (3D) space based on three orthogonal planes (TOP). The HOOF characterizes the motion information by inducing a very characteristic optical flow profile to distinguish smoky objects and non-smoky objects in successive ROIs based on the fact that the smoke is ejected from vehicle exhaust port and then gradually spreads around. The above three groups of features are complementary, and we fuse them to increase algorithm robustness. Experiment results show that our method achieves better performances than existing methods.