改进食肉植物算法优化的车辆长期目标跟踪器

研究了群体智能算法与目标跟踪相关的机理，提取了跟踪区域快速方向梯度直方图(FHOG)作为特征模板，利用食肉植物算法在图像搜索区域搜寻目标位置，基于巴氏距离作为模板匹配的相似度函数设计了一种食肉植物算法跟踪框架；考虑到实际跟踪过程的遮挡、背景杂乱等复杂状况，设计了一种短期记忆模块来预测跟踪过程中食肉植物算法的初始化个体，该模块采用高斯分布方式，利用视频序列前2帧中目标位置来预测运动轨迹；为更好地利用食肉植物算法优化目标跟踪，在迭代过程中设计了一种随机跟随策略与种群划分机制，并作为搜索策略引入跟踪框架中；为弥补方向梯度直方图单特征对目标的表述能力不强的问题，在FHOG基础上融合了ResNet-50中的Conv2-1与Conv4-1两层特征，并在融合特征的基础上设计了一种自适应学习率动态更新模板；在种群维度上添加二维比例感知因子，使得目标窗口的长宽比各自变化，更好适应目标窗口的尺度变化。研究结果表明: 随机跟随策略和种群划分机制的引入显著改善了食肉植物算法的迭代速度和寻优能力；融合特征和自适应模板更新增强了目标特征表述能力, 解决了由于遮挡等情况学习到大量无效信息，从而导致特征模板退化的问题；算法在跟踪UAV123中几个具有挑战性的车辆视频序列时性能优越，精确度和成功率分别为0.81和0.58，速度为每秒11.05帧；与同类算法相比，跟踪精度和鲁棒性均有大幅提升，能够适应复杂场景中的环境变化，对目标车辆长期保持稳定跟踪。

Vehicle long-term target tracker optimized by improved carnivorous plant algorithm

The mechanism of the swarm intelligence (SI) algorithm related to target tracking was studied. The fast histogram of oriented gradients (FHOG) of the tracking region was extracted as a feature template, and a carnivorous plant algorithm (CPA) was employed to search for the target's position in the image search region. A tracking framework based on the CPA was designed using the Bhattacharyya distance as the similarity function for template matching. Considering the complexity situations in the actual tracking process, such as the occlusion and busy background, a short-term memory module was designed to predict the individual initialized by the CPA during the tracking. This module, utilizing a Gaussian distribution, predicted the motion trajectory according to the target's position in the first two frames of the video sequence. To better optimize the target tracking with the CPA, a random tracking strategy and a population division mechanism were developed in the iterative process and integrated into the tracking framework as a search strategy. To make up for the poor representation ability of the single feature of the target by the FHOG, Conv2-1 and Conv4-1 features of ResNet-50 were integrated on the basis of the FHOG. A dynamic update template of the adaptive learning rate was designed based on this fused feature. A two-dimensional scale perception factor was added to the population dimension, allowing the aspect ratio of the target window to vary, so as to better adapt to the change in the scale of the target window. Analysis results show that the introduction of the random tracking strategy and population division mechanism significantly improves the iteration speed and optimization capability of the CPA. The fused feature and adaptive template update enhance the representation of target features, addressing the issues related to learning abundant irrelevant information due to the occlusion and preventing feature template degradation. The proposed algorithm demonstrates notable performance in tracking challenging vehicle video sequences from UAV123. The precision and success rate are 0.81 and 0.58, respectively, with a speed of 11.05 frames per second. Compared with the algorithms in similar literature, the tracking accuracy and robustness of the proposed algorithm improves substantially, making it adaptable to environmental change in complex scenes and ensuring a stable long-term tracking of target vehicles.