针对坦克目标,采用形状上下文特征级的异源图像配准以及小波变换图像融合方法。在此基础上,通过提取融合图像HOG特征并生成SVM分类器模型实现坦克目标识别。构建了嵌入式的红外与可见光图像融合系统,定制了ARM+FPGA样机。采用图像行缓存、窗口缓存等方法对图像进行处理,实现加速遍历图像像素卷积运算。定制了融合算法的IP核,通过软硬件协同设计进行资源优化与合理配置。对复杂背景下的静止和运动坦克目标在PC与嵌入式平台进行了融合和识别实验,结果表明:系统平均耗时为51ms,帧率约为20Hz,满足目标检测识别的实时性要求。

According to tank target, the image registration based on shape context feature leveland wavelet transform image fusion methodis adopted. Subsequently, the target recognition process for tank is developed by extracting HOG features of the fusion image and generatingSVM classifier model. An embedded system infrared and visible light image fusion system is constructed, and an ARM/FPGA prototype iscustomized. The improved methods, such as image line buffering, window buffering methods, are used to process the image to acceleratethe convolution operation of traversing image pixels. The IP core of the fusion algorithm is also customized, and the performance of resourceis optimized and reconfigured, which is based on the software and hardware cooperative design process. Depending on fusion experimentsfor stationary and moving tank targets in PC or embedded system, comparisons results show that the average processing time of single imageis about 51 ms, the frame rate is about 20 Hz, which could meet the demands of real running time.