Registration-based Moving Vehicle Detection for Low-altitude Urban Traffic Surveillance

Vision based vehicle detection attracts much attention in low-altitude platform for urban traffic surveillance which has great potential practical value. The major difficulties of moving vehicle detection are two-fold: 1) the unconstrained motion of the camera platform; 2) crowded vehicles are connected and hard to be segmented. To address these problems, we proposed an improved registration method to eliminate the camera platform motion, thus techniques in stationary vision surveillance such as frame subtraction can perform well on the registered video to detect moving vehicles rapidly and robustly. For registration, control points which should be in motionless regions play an important role; however two consequent images cannot be consistently transformed to the same affine coordinates, because some control points fall into the moving vehicle regions which cover a large area in the camera view. Compared with existing methods, adaptive sized windows are used to efficiently select control points efficiently only in static regions; and then frame subtraction is performed to segment moving regions and moving vehicles can be detected by using appearance based classification. Experimental results show that our method has better overall performance (i.e. higher detection rate, lower false positive rate and higher detection speed) and is promising for real-time application.