Robust image segmentation for overhead real time motorbike counting

Motorbikes are often difficult to detect in overhead road traffic images due to the variability of color, size, shape as well as trajectories. This paper tackles the problem of robust and real time image segmentation for motorbike counting. First of all, we perform background subtraction. Foreground blobs are then refined with Laplacian densities. This fusion enables to achieve a significant robustness to cast shadows. Thus, simple features, such as area, height and width, can be used to discriminate motorbikes from other vehicles. Our real time algorithm achieves interesting performances on multiple real traffic video sequences.