基于跟踪时序的多阶段目标融合算法OA

This paper presents a multi-stage object fusion strategy based on tracking time series to enhance object perception based on multi-sensor fusion in rail transit scenarios.Individual-sensor object detection and tracking are first carried out using camera and LiDAR;a multi-stage object fusion algorithm is then used based on tracking time series data to address omissions from the individual detection stage,thus increasing object perception performance.The algorithm employs a joint similarity calculation approach based on point-box displacement and the normalized Gaussian Wasserstein distance(NWD)to improve fusion accuracy.Additionally,the incorporation of an expanded Kuhn-Munkres(KM)algorithm enables matching a single visual detection to LiDAR clustering to further increase fusion accuracy.Experiments with perception equipment installed on a train show that this method produces real-time fusion results for objects ahead of the train,achieving fusion accuracies up to 95.1%for trains and 77.7%for pedestrians,with an average time consumption of 34.79 ms,providing reliable warnings to enhance active safety for trains.