目标检测通常利用复杂的、高维度的特征来提高其检测精度, 而高维特征往往会产生较高的计算复杂度和存储开销。经典的特征压缩算法常常被用于目标检测系统以实现特征降维, 但在其求解过程中会涉及到大量的矩阵分解运算, 从而降低了算法的实时性。针对此问题, 提出一种基于随机映射的特征压缩算法。该算法仅通过一个稀疏随机矩阵和简单的矩阵乘法运算就实现了特征从高维空间到低维空间的映射。利用经该算法压缩后的特征向量构建了Ada-Boost分类器, 实验结果表明, 该分类器在保证检测精度的前提下, 提高了目标检测的实时性。

The complicated and high-dimensionality features are always employed in object detection for improving the detection accuracy. Higher dimensionality features often yield higher computational complexity and memory cost. Therefore, feature compression algorithms are applied for reducing the dimensionality. The classical feature compression algorithms such as principal component analysis (PCA) and singular value decomposition (SVD) involve a large number of matrix decomposition operations, which are inefficient. To address this problem, a feature compression algorithm based on random projection was proposed. In this algorithm, the high-dimensionality feature was mapped into the low-dimensionality feature space by a sparse random matrix and efficient matrix multiplication. The compressed feature vectors were applied to build an Ada-Boost classifier, and the experimental results show that the detector not only guarantees the detection accuracy but also improves the detection speed.