摘要：

传统的目标检测算法通常预设许多候选的目标边界框(“锚框”)穷举待检测目标的潜在位置, 在对边界框进行置信度计算后筛除冗余边界框确定人物位置, 此类检测方法需要复杂的后处理程序, 检测效率低。针对待检测的人物目标, 提出基于热力图预测的免“锚框”目标检测算法, 将人物目标的检测转化为对人物热力图的最大值, 即目标中心点的检测。通过中心点进行目标尺寸的回归, 最终确定目标位置, 免除对“锚框”的依赖和计算, 从而有效降低计算成本, 大幅提高目标检测的速度。实验结果表明: 与传统基于“锚框”的检测算法Faster R-CNN和SSD相比, 所提算法目标检测速度大幅提升, 达到45帧/s, 同时检测精度相较前两者在不同数据集上均有所改善。在现实场景测试中, 即使视频中存在人物交叉遮挡情况, 该算法也能实时跟踪和精准检测人物位置, 达到实时检测的目的。

关键词: 目标检测, 免“锚框”, 热力图, 中心点检测, hourglass网络

Abstract:

Traditional target detection algorithms usually preset a number of candidate target bounding boxes(referred to as anchor boxes) to exhaustively cover potential target locations. Afterward, they filter out redundant bounding boxes to determine the target's location by calculating the confidence scores of the remaining boxes. This detection method requires complex post-processing and has low detection efficiency. To address these issues, this study proposes an anchor-free target detection algorithm based on heat map prediction. The algorithm transforms the detection of the character target into the detection of the maximum value on a character heat map, identifying the target's center point. By regressing the target size from the center point, the final target position can be determined, eliminating the need for anchor boxes and their associated calculations. This approach effectively reduces computational costs and significantly improves detection speed. Experimental results show that, compared to traditional anchor-based detection algorithms like Faster R-CNN and SSD, the proposed algorithm increases target detection speed to 45 frame/s. Additionally, detection accuracy on various datasets is significantly improved over the previous methods. In realistic scenes tests, the algorithm can track and accurately detect character positions in real-time, even in scenarios with character occlusions, achieving the goal of real-time detection.

Key words: target detection, anchor-free, heat map, center point detection, hourglass network