Design of automatic target recognition and detection system for sensors based on improved machine vision control

The aim of this study is to design and implement a sensor-based automatic target recognition and detection system with machine vision control. The system achieves high-precision detection of targets in complex environments by integrating multiple sensors, including industrial-grade color and infrared cameras, VelodyneHDL-64E lidar, ultrasonic arrays, and high-quality IMU devices. Through multi-sensor data fusion, preprocessing techniques, and feature extraction methods, the experimental results show that the system is able to achieve high-precision target detection in different scenarios. FasterR-CNN and its improved version of the model perform well in the experiments, especially after the introduction of the feature pyramid network (FPN) and the attention mechanism, which significantly improves the detection rate and the overall performance of the small targets. Experimental results show that the multi-sensor fusion system significantly improves the performance in target detection, with the accuracy of RGB cameras increasing from 85% to 92% and the recall rate increasing from 78% to 88%. After introducing the feature pyramid network (FPN) and attention mechanism, the detection accuracy of the Faster R-CNN model for small targets increased from 70% to 75%. Although the processing speed decreased slightly (from 20fps to 15fps), the overall detection accuracy and robustness were significantly enhanced. In addition, the model pruning technology increased the processing speed to 12fps while maintaining high accuracy, which is suitable for real-time applications. The model pruning technique successfully realizes the lightweighting of the model while maintaining high detection accuracy, which provides the possibility of real-time target detection for embedded devices.