High-Accuracy Dynamic Target Detection Through Low-Readout Channel Thermopile Infrared Array Detector Enhanced by Machine Learning Techniques

Infrared (IR) detector, due to its low cost, high sensitivity, and broad spectral response, has been extensively utilized in real-time dynamic target detection. In this study, a noncontact dynamic detection system, combining microelectromechanical systems (MEMS)
$4\times 8$
thermopile IR array detector with machine learning techniques was proposed, achieving real-time and high-accuracy target detection. Through low-readout channel system design, only four independent signal outputs were adopted. Each channel consisted of eight serially connected units, which utilized a double-layer radially distributed thermopile structure. This simple system significantly enhanced the output voltage of the detector while simultaneously reducing the complexity and cost of the subsequent circuitry and signal processing. We achieved a short response time of 13.2 ms and a high responsivity of 75.2 V/W at room temperature. Furthermore, continuous IR signals were converted into images containing temporal information, which were then processed through a convolutional neural network (CNN). The experimental results indicated the system can successfully recognize three different motion gestures with an accuracy of up to 97.92%. This exceptional dynamic target detection system holds promising prospects for applications in autonomous driving and human–machine interaction.