目标的检测与跟踪一直是监控领域的研究热点。同时，目标检测的设备多种多样，包括有源雷达、被动雷达网络、双目视觉等。因此，是否能在监测区域同时利用多种探测平台对区域进行监测从而得到更好的检测效果？本作目的即在于设计一种处理框架使得多种检测平台所得数据可以统一处理。虽然针对各个检测平台与数据融合处理的算法已有很多，但时在统一处理框架中检测得到目标轨迹的效果依然难以令人满意。无论是单个平台的检测方法还是数据融合的目标跟踪算法均存在提升空间，针对各个检测平台与数据融合方法的不足，提出了一系列改进算法。

1. 针对有源雷达目标跟踪中关于检测性能、虚警率、实时处理、存储空间限制等问题。本文首先提出了一种基于采样和时空恒虚警杂波图的背景提取算法。该算法将检测过程分为两部分，粗检测与精检测，基于采样的粗检测通过采样仅对目标存在的大致区域进行检测从而保障了算法的实时性与较低的内存需求。在精检测阶段中，疑似目标存在区域被分为三类，单目标、多目标与杂波区域，并利用不同模型进行估计。其次，利用所提基于多轮廓跟踪的目标检测算法对单目标区域进行精检测，通过多个检测门限下的目标轮廓与区域生长所得目标区域对疑似目标区域进行目标似然估计。由于对多个检测门限的使用，该方法在不同噪声基底中对目标进行检测具有较好的检测效果。

针对传统处理方法在无源雷达网络中检测概率较低，目标数量多时计算量较大的问题，提出了基于波形投票的目标成像算法。在每一个采样时间，雷达网络的每个节点均可得到回波。然后利用所提无源雷达网络成像算法得到监视区域的像。相比传统算法中对单个节点使用CFAR进行目标检测，该方法保留了更多的原始信息，保障了系统对弱小目标的检测能力。其次，利用所提基于罗尔定理的图像分割算法对所得监视区域的图像进行检测，得到疑似目标存在的区域。计算每个疑似区域的目标似然概率，并通过一个较低的检测门限进行检测，从而提升弱小目标检测能力。

针对双目视觉系统中，现有立体匹配算法均利用像素本身以及其周边纹理特征进行匹配，算法计算量大，效果不理想的问题，提出了一种基于超像素的立体匹配方法。将局部区域内的相似像素聚类为超像素。然后利用超像素的内部相似性和外部相似性来匹配不同图像中的超像素。该方法可通过更低的计算量获得较低的区域匹配错误率。在得到多路双目视觉系统的点云之后，搜索区域的目标可以通过多路点云被检测出。提出了一种基于三维区域生长的目标检测算法将同一目标的点分为一簇。同时计算每簇点云的目标似然概率作为该簇点云的权重。

针对传统多传感器的目标跟踪系统在杂波较多，目标回波较弱时存在跟踪效果差，虚警点较多的问题。通过将目标扩展、目标机动、密集目标、弱小目标和强杂波背景问题综合考虑，利用检测前跟踪的思想提出了基于多平台目标跟踪系统框架。该框架可以利用多个平台所得点迹在密集杂波中对密集、机动、弱小目标进行有效检测。该方法中利用检测前跟踪思想实现多个平台点迹的联合处理。提出基于三维映射的检测前跟踪算法来解决目标扩展、弱小目标密集和杂波等问题。各个平台点迹放在一起按照时间分为时间子窗，若干子窗又交叉组成检测所用时间窗。对每个时间窗中点迹利用所提基于三维映射的检测前跟踪算法得到该时间窗之内的轨迹段，而此阶段的结果即为每个时间窗内的轨迹片段。提出的基于狮群繁殖模型的轨迹段关联算法用于解决密集目标与机动目标的跟踪问题。算法迭代中，普通轨迹段仅与分数较高的轨迹段进行关联，以减少计算两轨迹段之间距离的次数。可在较低的计算量下获得完整的目标轨迹，同时也可在计算量恒定的情况下得到更优的轨迹。

Target detection and tracking has always been a research hotspot in the field of monitoring. At the same time, target detection equipment is various, including active radar, passive radar network, binocular vision and so on. Therefore, can a variety of detection platforms be used to monitor the area at the same time in order to get a better detection effect?The purpose of this project is to design a processing framework so that the data from various testing platforms can be processed uniformly.Although there are many algorithms for each detection platform and data fusion processing, the detection effect of target trajectory in the unified processing framework is still not satisfactory.There is room for improvement in both the detection method of a single platform and the target tracking algorithm of data fusion. In view of the shortcomings of each detection platform and data fusion method, a series of improved algorithms are proposed.

1. Aiming at the problems of detection performance, false alarm rate, real-time processing and storage space limitation in active radar target tracking. Firstly, a novel detection method which bases its principle on sampling and spatio-temporal detection is proposed. The method consists of two stages, coarse detection and fine detection. Sampling based coarse detection is designed to guarantee the real-time processing, low memory requirement by locating the area where targets may exist in advance. In the stage of fine detection, the suspected target area is divided into three categories: single-target, multi-target and clutter area, and different models are used for estimation. Secondly, the proposed target detection algorithm based on multi-contour tracking is used to accurately detect the single target region, and target likelihood estimation is carried out for the suspected target region through the target contour under multiple detection threshold and the target region derived from the growth of the region.Because of the use of multiple detection thresholds, this method has a good detection effect in detecting targets in different noise bases.

2. Aiming at the problem that the detection probability is low in the passive radar network with the traditional processing method and the computation is large when the number of targets is large, a target imaging algorithm based on waveform voting is proposed. At each sampling time, the echo can be obtained at each node of the radar network.Then the passive radar network imaging algorithm is used to get the image of the surveillance area.Compared with the traditional algorithm that USES CFAR for target detection of a single node, this method retains more original information and ensures the system's ability to detect dim small targets.Secondly, the proposed image segmentation algorithm based on Rolle's theorem is used to detect the image of the obtained monitoring area, and the area where the suspected target exists is obtained.The target likelihood probability of each suspected area is calculated, and a low detection threshold is used for detection, so as to improve the detection ability of dim small targets. 

3. In binocular vision system, the existing stereo matching algorithms all use the pixel itself and its surrounding texture features to match, the algorithm computation is large, and the effect is not ideal, a superpixel based stereo matching method is proposed. The similar pixels in the local area are clustered as a superpixel. Then the internal similarity and external similarity of the superpixels are used to match the superpixels in different images. This method can obtain lower region matching error rate with lower computation effort.After the point cloud of the multiplex binocular vision system is obtained, the target of the search area can be detected by the multiplex point cloud.A target detection algorithm based on three-dimensional region growth is proposed to divide the points of the same target into a cluster.At the same time, the target likelihood probability of each cluster cloud is calculated as the weight of the cluster cloud.

4. For the traditional multi-sensor target tracking system, the tracking effect is poor and there are many false alarm points when the target echoes are weak. By comprehensively considering the problems of target expansion, target maneuver, dense target, weak target and strong clutter background, a multi-platform target tracking system framework is proposed based on the idea of pre-detection tracking.This framework can effectively detect dense, mobile and small targets in dense clutter by using plots obtained from multiple platforms. In this method, the idea of TBD is used to realize the joint processing of plots on multiple platforms. A 3-dimensional projection based TBD algorithm is proposed to solve the problems of target expansion, small and weak target density and clutter.The points of each platform are put together and divided into time sub-windows according to the time. Some sub-windows cross to form the time window for detection. The 3-dimensional projection based TBD algorithm is used to obtain the track segment in each time window, and the result of this stage is the track segment in each time window. The trajectory segment correlation algorithm based on the breeding model of lions is proposed to solve the tracking problem of dense targets and maneuvering targets. In the algorithm iteration, the ordinary track segment is only associated with the track segment with higher score, so as to reduce the number of times of calculating the distance between the two track segments. The complete target trajectory can be obtained at a lower computational cost, and the better trajectory can be obtained at a constant computational cost.