Spatial Data Association Target Fusion of Multi-sensor for Vehicle Detection

This paper proposes a target detection network based on three-dimensional decision fusion framework, which can overcome data bias and achieve more accurate data asso-ciation. Combined with cost volume and regression network, the distance information is predicted. The distance estimation of targets is incorporated into the bounding boxes as the results of the image detection branch. By inputting multi-dimensional detection bounding boxes of sensors, the proposed network can more accurately realize data associations. The detection results from multi-sensor are input into the proposed model, ensuring maximum depth analysis of multi-sensor spatial information. This fusion model thoroughly breaks the complexity of traditional data correlation, and adaptively fuses detection results with biases among multi-sensors. Through the adaptive and fitting ability of neural network, it effectively overcomes the deterministic limitations imposed by threshold settings in traditional fusion methods, making the fusion of detection results simpler and more accurate. Finally, the algorithm is validated in Carla, and the results show that the proposed model achieves the single-frame average detection accuracy of 95 % in spatial level within the range of 0-56m.