Monocular Non-linear Photometric Transformation Visual Odometry Based on Direct Sparse Odometry

When no photometric calibration is available in Direct Sparse Odemetry (DSO), a direct sparse odometry with method of Monocular Non-linear Photometric Transformation Visual Odometry is proposed to improve tracking accuracy and robustness of DSO. Compared with DSO method, a new regularization item and a non-linear brightness transfer function are integrated in proposed method, so it can better simulate the photometric calibration process. The proposed method is evaluated on 50 different sequences of TUM MonoVO datasets containing several hours of video in total. The results show that the rms of aligned error of the proposed Monocular Non-linear Photometric Transformation Visual Odometry method is 11.4% lower than original Direct Sparse Odometry on average. Successful running times of the proposed method are 4.4% more than original Direct Sparse Odometry when the environment light source intensity changing.