人体检测与异常行为识别联合算法

近年来，异常行为识别算法取得了一定的研究进展，但是针对复杂环境、人体遮挡、动作相似度高等多种挑战，识别算法的适应性、效率、准确性都有待进一步提高。为了解决以上问题，提出了基于特征增强的人体检测与异常行为识别联合算法，首先将视频序列分别送入人体检测网络和特征加强网络，再采用爱因斯坦求和法将特征加强网络输出的多头卷积注意力特征与人体检测网络输出的热力图特征融合，得到加强融合特征，然后利用检测网络输出的人体目标位置特征信息和ROI Align模块对加强融合特征进行人体ROI(region of interest)区域特征截取，得到人体ROI区域加强融合特征，最后将人体ROI区域加强融合特征送入Transformer时序建模网络模块进行人体行为特征时序建模和识别。所提算法充分利用检测网络中间过程产生的行为主体区域特征，弱化了复杂环境中背景的干扰，同时实现了检测网络的输出特征共享，避免了识别网络的二次特征提取过程，从而提高了网络运行效率，且利用Transformer网络的建模优势，能够充分挖掘人体行为空间特征、时序特征以及之间的跨域特征的优势。实验结果表明：所提算法在提高了网络效率的同时大幅度地...

A Join Algorithm of Human Detection and Abnormal Behavior Recognition

In recent years, the abnormal behavior recognition algorithm has achieved good research progresses. However, the adaptability, efficiency and accuracy of the recognition algorithm need to be further improved for the challenges of complex environment, human occlusion and high action similarity. In order to solve the above problem, a joint algorithm of human detection and abnormal behavior recognition based on feature enhancement is proposed. First, the video sequence is sent to the human detection network and the feature enhancement network respectively, and then the features of multi-head convolutional attention output by the feature enhancement network and the heat map feature output by the human detection network are fused by the Einstein summation method to obtain enhanced fusion feature. Then, the human body target position feature information output by the detection network and the ROI Align module are used to intercept the enhanced fusion features. Finally, the enhanced fusion features of the human ROI region are sent to the Transformer time series modeling network for temporal modeling and recognition of human behavioral. The advantages of the proposed algorithm is that it makes full use of the main area features generated in the intermediate process of the network, avoids the secondary feature extraction process of the identification network to improve the network operation efficiency, and uses the Transformer time series modeling network to fully exploit the advantages of cross-domain features between human behavior space and time series features. The experimental results show that the proposed algorithm not only improves the efficiency of the network, but also greatly improves the recognition accuracy of the network, and achieves the expected effect.