Improved Lightweight Head Detection Based on GhostNet-SSD

This abstract proposes an algorithm for human head detection in elevator cabins that addresses the challenges of improving detection accuracy, reducing detection speed, and decreasing the number of parameters. The algorithm is based on GhostNet-SSD and includes several improvements, such as an efficient coordinate attention mechanism to replace the Squeeze-and-Excitation attention mechanism, optimization of auxiliary convolutional layer with large parameter weight, and adjustment of anchor ratio based on the statistical results of human head labeling frame. In addition, data normalization and convolutional fusion methods are used for inference acceleration. The algorithm was tested on JETSON XAVIER NX development board and achieved a new state-of-the-art 97.91% AP at 61FPS, outperforming other detectors with similar inference speed. The effectiveness of each component was validated through careful experimentation.