A stacked meta approach for object detection to reduce false positives in highly complex videos

False positives are a major problem when deploying object detection models in real-world conditions. Highly complex scenes are particularly difficult to process by standard object detection models. A novel meta-approach of stacked detection and the use of multiple frames to evaluate the preliminary detections is presented. The stacked approach leverages different types of architectures and performs multiple detections to reduce the number of false positives. The approach was qualitatively validated with videos taken from construction sites and compared with some of the most used architectures, i.e., Faster-RCNN and RetinaNet. Our approach can reduce the number of false positives and increase the detection accuracy.