为了提高末敏弹在复杂背景条件下对装甲目标的识别能力, 将线阵激光雷达作为探测器, 结合卷积神经网络对线阵激光雷达距离像进行目标分类与识别.利用末敏弹边旋转边下降的运动特点, 实现对扫描区域的距离成像, 并通过采样率控制及插值等算法将原始距离像构造成适用于卷积神经网络的灰度像.针对弹载高实时性、小体积和低功耗的要求, 建立了由两层卷积层和一层全链接层构成的浅层卷积网络, 选用Xilinx ZYNQSoC 芯片作为硬件平台, 通过基于HLS技术和SDSoC开发环境将卷积操作放在端进行硬件并行加速.缩比模拟试验结果验证了该方法具有较高的目标识别精度, 对复杂背景下的装甲目标也能有效识别.ZYNQSoC的PL硬件相较于普通CPU方案, 加速性能提升了5倍, 能够满足弹载的要求.

In order to improve the detection ability of the terminal sensitive projectile to the armored target under complex background conditions, the linear array LiDAR is used as the detector and the convolutional neural network is combined to classify and identify the range profile of the linear array LiDAR. The range imaging of the scanning area is realized by using the steady-state motion characteristics of the terminal sensitive projectile dropping while rotating. The original range profile is constructed into a range profile suitable for convolutional neural network by sampling rate control and interpolation. The convolutional neural network consisting of two convolutional layers and one full link layer are established to meet the requirements of high real-time, small size and low power consumption on missile-borne. Xilinx ZYNQSoC chip is selected as hardware platform, and hardware acceleration is achieved by placing convolution operation on PL end of ZYNQSoC based on HLS technology and SDSoC development environment. The scaled simulation experiment proves that the method has high target recognition accuracy and can effectively recognize armored targets in complex background. Through placing convolution operation on PL end of ZYNQSoC, the performance acceleration is five times faster than that of CPU, which can meet the requirements of missile-borne.