结合语义分割和特征融合的行人检测方法

Pedestrian detection method combining semantic segmentation and feature fusion

针对传统行人检测方法在复杂场景下存在遮挡行人和小尺寸行人检测效果差的问题,提出一种结合语义分割和特征融合的行人检测方法。该方法的网络结构以区域全卷积神经网络为基础框架,根据行人检测任务进行改进。使用深度残差网络提取出多尺度的特征映射图;通过全卷积语义分割网络,得到对应的语义分割图;利用特征融合模块构造出融合特征图;将融合特征图送入区域建议网络和检测网络,完成行人检测。在Caltech和ETH行人检测数据集上进行试验验证,结果表明,与传统行人检测方法相比,该方法的行人检测准确率得到提高,同时检测速度满足实时性要求。

Aiming at the problem that traditional pedestrian detection methods had poor performance on pedestrians with small size and occlusion in complex scenes,a pedestrian detection method combining semantic segmentation and feature fusion was proposed.This method regarded region-based fully convolutional network as foundation framework,which was improved according to pedestrian detection task.The residual network was used to extract multi-scale feature maps.The corresponding semantic segmentation graph was obtained by means of full convolution semantic segmentation network.The feature fusion module was designed to construct an integrated feature map.The integrated feature map was sent into the regional proposal network and detection network to complete the detection of pedestrians.Experimental verification was completed on the Caltech and ETH pedestrian detection datasets.Experimental results demonstrate that the accuracy of pedestrian detection is improved significantly compared with those of some traditional detection algorithms.Meanwhile,the detection speed meets the real-time requirement.