Analysis of Missed Detections and False Alarms in Radar Occupancy Grid Map Vehicle Detection Networks

Millimeter-wave radar can collect data and form images under poor lighting conditions. Radar occupancy grid map vehicle detection plays an important role in confined disaster spaces. This paper uses a YOLO-based neural network model for vehicle object detection. The detection results show that false alarms and missed detections occur when there are objects with similar shapes and sizes near the vehicles or when the vehicles are parked irregularly. These issues can lead to misjudgments in specific situations such as fires in enclosed spaces, affecting rescue efficiency. This paper summarizes the causes of these problems. For the radar occupancy grid map vehicle dataset, Grad-CAM is used to visualize the feature extraction layers of the network model. It is further clarified that the cause of false alarms is the low distinguishability between feature encoding vectors, and the cause of missed detections is the low weight of vehicle features. A dynamic encoding approach is proposed to improve the distinguishability of feature vector encoding and to assign correct weights to the features extracted by the network model.