一种改进Mask R-CNN的运动目标定位方法

针对在复杂场景下传统算法对运动目标分类精度不高且无法直接输出关键点空间坐标这一问题，本文提出了一种基于改进Mask R-CNN的运动目标定位方法。该方法在TensorFlow 框架下采用自制数据集预训练提取多尺度的卷积，将采集到的网络视频流通过RoIAlign算法处理获得像素级的特征并匹配SIFT特征点建立空间坐标系统，在空间约束下结合Kalman filtering补偿运动轨迹，通过投影变换矩阵将像素坐标转换为世界坐标。实验表明该方法提高了模型在干扰背景下的鲁棒性，平均误差在8cm以内，实现了在复杂场景下可以获得实时高精度的定位信息。

AN IMPROVED MOVING TARGET LOCALIZATION METHOD FOR MASK R-CNN

Aiming at the problem that traditional algorithms do not classify moving targets accurately in complex scenes and cannot directly output the spatial coordinates of key points, a moving target localization method based on improved Mask R-CNN is proposed. Under the TensorFlow framework, a self-made dataset is used to pre-train to extract multi-scale convolutions. The collected network video stream is processed through the RoIAlign algorithm to obtain pixel-level features and match SIFT feature points to establish a spatial coordinate system. In combination with Kalman filtering under spatial constraints Compensate the motion trajectory and transform the pixel coordinates into world coordinates through a projection transformation matrix. Experiments show that the robustness of the model is improved under the interference background, and the average error is within 8cm, which realizes that real-time high-precision positioning information can be obtained in complex scenarios.