Enhancing Doppler Ego-Motion Estimation: A Temporally Weighted Approach to RANSAC

Ego-motion estimation from millimetre-wave Doppler are subjected to high levels of outliers caused by multipath reflections. A common method to reduce the effect of outliers is to fit measurement points to a velocity model using the random sample consensus (RANSAC) algorithm, where the velocity model is traditionally parameterized from data collected from a single time step. In this paper, we show that the temporal relationship of data points between successive measurements can be exploited to improve velocity estimation. Two variations of RANSAC with an integrated weighted sliding window are proposed. Each point in the window receives a weight that decreases over time. In the first algorithm, points are sampled with a temporally weighted probability and the velocity model is generated using least square regression (LSQ). In the second algorithm, points are sampled with a uniform probability but the model is parameterized with a temporally weighted LSQ. The motion of the platform is then calculated for both methods. Experimental results using data from three indoor locations demonstrate an average position accuracy improvement of 19.5% over conventional RANSAC implementations.