融合视觉与雷达数据的改进粒子滤波车辆目标跟踪

针对视觉跟踪中由于尺寸变化累积误差导致目标丢失的问题，提出一种融合视觉与毫米波雷达数据的改进粒子滤波车辆跟踪算法。首先，引入遗传算法改善标准粒子滤波中的粒子退化与粒子衰退问题，根据退化程度计算动态自适应的遗传交叉概率，并利用高斯分布替代平均分布计算种群适应度。然后，将图像HSV直方图特征与改进粒子滤波算法结合，实现车辆多目标跟踪。最后，通过雷达目标投影点与视觉跟踪框的位置关系实现关联匹配，利用深度信息修正跟踪框的位置与尺寸。实验结果表明，相对于标准粒子滤波，改进的粒子滤波算法可以使平均跟踪准确率与精度分别提高22.1%与21.1%。相对于仅采用视觉跟踪，融合雷达数据的跟踪算法能够使车辆目标跟踪精度再次提高9.2%。

Improved particle filter vehicle tracking based on vision and radar sensor fusion

In order to solve the problem of target loss due to size change in visual tracking, an improved particle filter vehicle tracking algorithm fusing visual and millimeter wave radar is proposed. First, genetic algorithm is used to improve particle degradation and resampling induced by standard particle filtering. The dynamic adaptive genetic cross probability is calculated according to particle degradation degree, and instead of mean distribution Gaussian function is used to calculate fitness. Then, the HSV histogram features are combined with the improved particle to achieve vehicle multi-target tracking. Finally, the location and size of the tracking bounding boxes are modified by the range information from radar. The experimental results show that compared with the standard particle filter, the improved particle filter algorithm significantly improves the multi-object tracking accuracy (MOTA) and multi-object tracking precision (MOTP) by 22.1% and 21.1% respectively. Compared to visual tracking algorithms, the tracking algorithm that fused radar data can improve the precision by 9.2% again.