面向实际应用场景的实时行人检测算法研究

Research on Real-Time Pedestrian Detection Algorithm for Practical Application Scenarios

为提高基于计算机视觉的行人检测方法的实时精确度,克服实际应用场景下存在的背景复杂性、行人姿势多样性和行人遮挡等不利因素的影响,提出一种改进的行人检测算法MCDET。新方法使用改进的M2Det网络,将原有的模块从8个降低到4个,减少运算参数,提升速度;使用VGG作为特征提取网络,并将浅层特征与深层特征进行融合,添加CBAM注意力机制,以提高整体精度。经实验证明本方法拥有较高的精度和较快的速度,具有良好的泛化性,满足实时性要求。

In order to improve the real-time accuracy of pedestrian detection method based on computer vision and overcome the unfavorable factors such as background complexity,pedestrian posture diversity and pedestrian occlusion,an improved pedestrian detection algorithm MCDET is proposed.Using the improved M2Det network,the original modules are reduced from 8 to 4,so that the operation parameters are reduced and the speed is increased.VGG is used as feature extraction network,and shallow features and deep features are fused,and CBAM attention mechanism is added to improve the overall accuracy.Experiments show that the method has higher accuracy and faster speed,has good generalization and meets the real-time requirements.