Energy-efficient illumination-invariant change detection in DCT coefficient domain for vehicular black box camera

Abstract

An energy-efficient and highly accurate change detection method for a vehicular black box system by utilising information in a compressed domain is presented. Studies have shown that most automobiles spend a substantial percentage of time parked rather than driven. Thus, minimising energy consumption in a parked mode is crucial for vehicular black box systems design to prevent the battery being fully discharged. To meet a low energy requirement in parked mode, a hybrid video encoding engine with JPEG for parked and H.264 for driving mode is used. A change detection method is devised for parked mode using the JPEG encoder. The proposed change detection method detects events by utilising discrete cosine transform (DCT) coefficients that are readily available from the JPEG encoder at no extra computing cost. By using the ratio AC coefficients to a DC coefficient in each DCT block, illumination-resistant change detection has been successfully achieved. The illumination-invariant detection reduced the false positives while maintaining a very low false negative ratio. Experimental results demonstrate that the proposed method achieves 25.6-24.2% lower energy consumption and 25.4-1.2% better accuracy in illumination-varying circumstances compared with existing illumination-independent change detection algorithms.