【作者】 王鹏；

【导师】 王海燕；

【作者基本信息】 西北工业大学 ， 兵器科学与技术， 2022， 博士

【摘要】 为了全天候地掌握入侵武装力量信息,必须实现对战场区域中出现的目标进行快速准确检测与稳定跟踪,然而由于战场环境复杂,如微光环境、非目标物体遮挡、目标远距离成像等场景都给战场信息感知带来了巨大挑战。本论文针对复杂战场环境下信息感知问题,开展基于深度学习和相关滤波的小目标检测与跟踪方法研究,突破夜间微光环境下高质量图像增强算法、资源受限硬件平台下对远距离成像的小目标检测算法、遮挡形变条件下提升目标定位与尺度估计精度的跟踪算法等瓶颈,建立一套适应复杂场景和作战平台的小目标检测与跟踪试验验证系统,提高陆战场中典型小目标的检测性能与跟踪精度。论文的主要研究内容及创新点如下:(1)针对微光环境下作战平台在执行小目标检测与跟踪任务时难以获得高质量图像信息的问题,提出一种一体式网络微光图像增强算法(LLAON)。首先将微光图像分割为若干超像素并估计噪声纹理水平,利用滤波器平滑得到基础层和细节层;其次采用噪声纹理水平指导基础层和细节层自适应组合获得去噪且纹理完整的图像;然后将去噪图像应用于对比度增强网络中,通过卷积与池化操作,提取微光图像的关键特征;最后通过由乘法层与若干加法层构成的清晰图像生成模块,增强微光图像对比度获得清晰图像。(2)针对实际复杂作战场景要求对小目标检测同时兼顾速度和精度的问题,提出了一种融合轻量化特征提取网络的小目标检测算法(GFaster R-CNN)。首先基于Faster R-CNN模型,通过Res Net结合FPN构建特征提取能力更优的Res Net-FPN骨干网络;其次采用Ghost卷积模块轻量化重构Res Net-FPN,保证特征信息尽量完整的情况下减少冗余操作,加快特征提取效率;然后通过残差部分衔接轻量级通道注意力机制ECA模块,在不降低维度的情况下完成各个特征信道间的信息交互,提升算法的检测精度;最后采用特征级知识蒸馏的方法提升轻量级骨干网络的特征提取能力与模型的泛化能力,在保证检测精度情况下提升目标检测效率。(3)针对目标在遮挡及相似物干扰等复杂情况下的定位不准确、跟踪精度低问题,提出了一种时空正则项与连续卷积算子交互的目标跟踪算法(CO-STRCF)。首先在连续卷积算子的滤波器函数中加入时空正则化,保证了滤波器模型在时间上的连续性与泛化能力;其次采用帕萨瓦尔公式将模型转化为频域,并通过预设共轭梯度算法求解滤波系数;然后利用通道可靠性分析滤波器质量,将不同的可靠性系数融合到对应位置的权重响应图中,通过设置阈值减少滤波器求解迭代次数提高计算速度;最后构建尺度滤波器并在频域求取闭合解完成目标尺度估计,避免了位置和尺度估计不能同时达到最优的问题,有效提升复杂场景下的目标定位精度和跟踪精度。(4)针对目标在遮挡、旋转及尺度形变等情况下跟踪算法的成功率低的问题,提出了一种基于三维正则项和可靠性判别机制的目标跟踪算法(TDRR-RDM)。首先基于三维正则项构成目标响应自适应滤波器模型,并在求解过程中持续更新理想目标响应函数;其次采用帕萨瓦尔公式将模型转化到频域,利用交替方向乘子法将模型分解为两个子函数并求解得到滤波器系数;然后通过融合多重通道可靠性系数建立置信度判别机制来评测特征通道响应图对跟踪目标的突出性,提高模型对目标的表达能力,提升定位准确度;最后将目标图像转换至对数极坐标系下训练尺度滤波器,提高在目标旋转以及较大形变时的尺度估计精度,从而提高了跟踪成功率。(5)在实验室环境和外场环境模拟搭建了复杂场景并开展了实景试验。试验得到的实际测试结果进一步验证了本文所提各个算法的性能,为下一步的软件系统集成并最终实现与传统作战平台的快速集成提供了重要技术支撑。更多还原

【Abstract】 In order to obtain the information of invading armed forces under all-weather scenarios,it is necessary to realize accurate detection and stable tracking in the battlefield area.However,due to the complex battlefield environment,such as low illumination light,non target object occlusion,target long-distance imaging and other scenes,it has brought great challenges to information sensing of battlefield.Aiming at the problem of information sensing under complex battlefield environment,this research carries out the theoretical study of small target detection and tracking method based on deep learning and correlation filtering.It mainly studies the LLAON image enhancement algorithm,small target detection algorithm to extract salient features deal with long-distance imaging,and tracking algorithm to improve the accuracy of target positioning and scale estimation under occlusion conditions.A small target detection and tracking verification system suitable for complex scenarios is established to improve the detection performance and tracking accuracy of typical small targets in the battlefield.The main results and innovations of this research are as follows:(1)To deal with the difficulty of obtaining high-quality image information when the combat platform performs small target detection and tracking tasks in low-light-level environment,an integrated network low-light-level image enhancement algorithm is proposed.Firstly,the low-light-level image is segmented into several super pixels and the noise texture level is estimated.Different filters are used to smooth the image to get basic layer and detail layer.Secondly,the noise texture level is used to guide the adaptive combination of base layer and detail layer to obtain a noise free and texture complete image.Then the noiseless image is applied to the integrated network,and the key features of low-light-level image are extracted through convolution and pooling operation.Finally,through the clear image generation module composed of multiplication layer and several addition layers,the low-light-level image is enhanced to obtain a clear image.(2)Aiming at the problem that the small target detection must take into account both speed and accuracy under the actual complex scene,a small target detection algorithm integrating lightweight feature extraction network is proposed.Firstly,Res Net-FPN backbone network with better feature extraction ability is constructed with Res Net and FPN.Secondly,Ghost convolution module is used for lightweight reconstruction of Res Net-FPN to reduce redundant operations and speed up feature extraction efficiency.Then,in order to improve the detection accuracy of the algorithm,the residual part is connected with ECA module to complete the information interaction between each characteristic channel without reducing the feature dimension.Finally,the feature level knowledge distillation method is used to improve the feature extraction ability and generalization ability of the lightweight backbone network,so as to restore the accuracy of the detection model while ensuring the detection efficiency.(3)To solve the problems of inaccurate location and low tracking accuracy under the interference of occlusion and similar objects,a target tracking algorithm,named CO-STRCF,based on the interaction of spatio-temporal regular term and continuous convolution operator is proposed.Firstly,spatio-temporal regularization is added into the filter function of continuous convolution to ensure the continuity and generalization ability of the filter model.Secondly,pashawar formula is used to transform the filter model into frequency domain,and the filter coefficients are solved by preset conjugate gradient algorithm.Then,the channel reliability is used to analyze the filter quality,and different reliability coefficients are fused into the weight response diagram of the corresponding position.By setting the threshold,the number of iterations times are reduced so as to the calculation speed will improved.Finally,the scale filter is constructed and the closed solution is obtained in the frequency domain to complete the target scale estimation,which avoids the position and scale estimation cannot be optimized at the same time.(4)Under the case of occlusion,rotation and scale deformation,a target tracking algorithm,named TDRR-RDM based on three-dimensional regular term and reliability discrimination mechanism is proposed.Firstly,a target response adaptive filter model is constructed by combining the spatial regularization term,time regularization term and target response regularization term.The ideal target response function is continuously updated during the solution process.Secondly,pashawar formula is used to transform the model into frequency domain,and the alternating direction multiplier method is used to decompose the filter model into two sub functions and solve them to obtain the filter coefficients.Then,the reliability coefficient is multiplied to improve the positioning accuracy.Finally,the target image is transformed into log polar coordinate system to train the scale filter,it can improve the scale estimation accuracy and the success rate of tracking when the target rotates and deformation occurs.(5)The complex scenes were simulated in the laboratory environment and outfield environment to carry out the real scene experiments.The experimental results are consistent with the simulations,which further verifies the effectiveness and feasibility of the algorithms proposed in this research.It provides an important technical support for the software system integration and the final realization of rapid assembly with the traditional combat platform.更多还原