Improving pedestrian safety using combined HOG and Haar partial detection in mobile systems

Objective: The objective of this study is to develop a novel algorithm on a mobile system that can warn drivers about the possibility of a collision with a pedestrian. The constraints of the algorithm are near-real-time detection speed and a good detection rate. Method: Histogram of gradients (HOG)-based detection is widely used in pedestrian safety applications; however, it has low detection speed for real-time systems. Hence, it has no direct usage for mobile systems. In order to achieve near-real-time detection speed, partial Haar transform predetections are applied to an image before HOG detection. The partial and HOG detections are merged and a score-based confidence level is defined for the final detection phase. In this way, the outcome is prioritized and different warning levels can be issued to warn the driver before a possible pedestrian collision. Results: The proposed algorithm provides an increase in detection speed (from 46 to 76 fps) and detection rate (from 80 to 91%) with respect to HOG-based pedestrian detection. It also improves confidence of the results by multidetection merging and score assignment to detections. Conclusions: Performance improvement of the algorithm is compared with respect to state-of-the-art detectors/algorithms. Based on the detection rate and detection speed performance, it can be concluded that the proposed algorithm is suitable to be used for mobile systems to warn drivers about the possibility of collision with a pedestrian.

研究目标：本研究旨在开发一款运行于移动系统的新型算法，用于向驾驶员预警与行人发生碰撞的可能性。该算法需满足两项约束条件：近实时检测速度与优异的检测率。
研究方法：基于方向梯度直方图（Histogram of Oriented Gradients, HOG）的检测方法已广泛应用于行人安全相关场景，但在实时系统中检测速度较慢，无法直接适配移动系统。为实现近实时检测速度，本研究在执行HOG检测前，先对图像进行局部哈尔变换预检测；随后将局部检测结果与HOG检测结果进行融合，并为最终检测阶段定义基于得分的置信度等级。通过该方案，可对检测结果进行优先级排序，并在潜在行人碰撞风险发生前，向驾驶员发出不同等级的预警。
实验结果：相较于传统基于HOG的行人检测算法，本文所提算法将检测速度从46fps提升至76fps，检测率从80%提升至91%。此外，通过多检测结果融合与检测得分赋值，该算法还提升了检测结果的置信度。
研究结论：本研究将所提算法的性能提升效果与当前主流检测器/算法进行了对比。基于检测率与检测速度两项性能指标，可得出结论：所提算法适配移动系统应用，能够实现对驾驶员的行人碰撞风险预警。