Complementary Object Detection: Improving Reliability of Object Candidates Using Redundant Detection Approaches

In actual advanced technical applications, like autonomous driving, Machine Learning is utilized. Most of these methods work well in certain and/or trained situations but can fail in unknown or uncertain situations. Therefore, overreliance might lead to safety-critical situations. Detecting objects appears as a key task for the safe operation of automated systems, like autonomous vehicles. To address potential failures of an object detection system, different redundant approaches can be used. Recent research aims for fusion, combining different modalities and architectures to utilize their advantages. It can be assumed that a combination of diverse approaches compensates each other's drawbacks and leads to improved reliability and robustness of the final prediction. In this contribution the fusion of detections of multiple detection systems at a detection level is studied using different opinion pooling strategies. The predicted detection score is calibrated using the true positive rate at a score level. This results in a standardized score over different detection approaches. Afterwards detection candidates of different approaches are associated and a new detection candidate is generated in a fusion stage. Therefore missed or false positive detections of one apporach can be compensated based on a redundent set of predicted object candidates. The aim is to highlight certain detections and to reduce the detection score of false positive detections. The fused approach generates a 3% improvement in comparison to the best individual results of single approaches, additionally improved robustness is achieved.

Keywords: Object detection, Decision fusion, Redundent systems, Opinion pooling, Autonomous driving, Robust object detection.

Download PDF