Objective With the rapid development of infrastructure industry in China, safety production issues in the construction field, represented by shield tunneling, are becoming increasingly prominent. Currently, construction site monitoring relies heavily on manual supervision, leading to low efficiency and difficulty in ensuring construction workers’ safety. Regarding the identity verification difficulties in facial recognition algorithms for non-front view faces caused by on-site helmet wearing violations, corresponding researches are conducted.

Method First, data preprocessing is performed. To address the image noise interference caused by vibration during shield tunneling process, the NLM (non-local means) algorithm is adopted for image denoising. To address the missing features of non-front view faces in construction scenarios, the traditional face recognition algorithm based on DREAM (deep residual equivariant mapping) is improved. By extracting the relationship between the face poses and face features, and considering the effect of face poses on feature correction from multiple dimensions, a non-front view face recognition algorithm based on multi-pose correction is proposed.

Result & Conclusion  The proposed algorithm can correct the extracted face features according to the face pose angles, thereby realizing the output face features pose-invariant. This solves the problem of multi-pose face image data and improves the accuracy of non-front view face recognition in shield tunnel construction scenarios.