融合回归网络和多尺度特征表示的实时行人检测

Real-time pedestrian detection based on regression network and multi-scale feature representation

提出了一种基于卷积神经网络(CNN)的实时鲁棒的行人检测算法。随着CNN在图像识别的广泛成功,CNN也被引用到行人检测领域,但这些方法还难以达到实时应用。在充分利用CNN强大特征表征能力的同时,提出了一种实时的行人检测算法,该算法将行人检测构建为一个回归问题,每张图像只需一次网络前向传播,直接输出行人包围框的位置以及置信度。此外,还提出了一种多尺度特征表示方法,进一步提高了行人检测的性能。最后在通用的Caltech行人检测数据集进行算法测试,实验结果验证了算法的有效性,平均检测速度可以达到50.6 fps,可以满足实时应用的需求。

A real-time and robust pedestrian detection method based on convolutional neural network( CNN) is proposed. Along with the success of CNN in image recognition,CNN has also been used in pedestrian detection,while these methods are far from the real-time requirements. On the assistance of the rich feature representation of CNN,a real-time pedestrian detection method is proposed,which formulates pedestrian detection as a regression problem. Every detection only needs a forward operation,and the location and confidence of pedestrian bounding boxes are output directly. Additionally,a multi-scale feature representation strategy is proposed,which further improves the detection performance. In the end,the method is tested on the Caltech pedestrian detection dataset,and the experimental results demonstrate the effectiveness of the proposed method,the average speed can achieve 50.6FPS,which can meet the real-time requirements.