Aiming at the problem of recognizing construction personnel’s safety gear in the context of smart construction site, this paper designs a surveillance video compression and target detection model based on convolutional neural network. The network is reconstructed by reversible convolutional blocks to save the memory consumption during training. ResNet-50 network is used as the feature extraction network of BCNN, and the task difference maximization mechanism is introduced to construct the RTF-BCNN target detection model. The performance level of video compression and target detection of the proposed model is examined through empirical analysis of power smart site monitoring videos. The proposed video compression model significantly improves the coding efficiency while maintaining the compression performance, with an average time saving of 25.446%, an average PSNR loss of 0.058 dB, and a bit rate increase of basically no more than 3%. Compared to the ResBCNN model without embedded TDM module, the RTF-BCNN model mAP accuracy is improved by 2.446%, the P-R curve area is larger, and the helmet AP increases from 91.89% to 98.26%. Compared with other model running speeds, the RTF-BCNN model FPS value reaches 67.842 frames/s, which is significantly higher than the comparison model. The experimental results show that the method designed in this paper can be effectively used to recognize protective and safety equipment and improve the safety management of construction site workers.
