Evaluation Methodology for Object Detection and Tracking in Bounding Box Based Perception Modules

Abstract

The aim of this work is to formulate a new metric to be used in the automotive industry for the evaluation process of software used to detect vehicles on video data. To achieve this goal, we have formulated a new concept for measuring the degree of matching between rectangles for industrial use. We propose new measure based on three sub-measures focused on the area of the rectangle, its shape, and distance. These sub-measures are merged into a General similarity measure to avoid problems with poor adaptability of the Jaccard index to practical issues of recognition. Additionally, we create method of calculation of detection quality in the sequence of video frames that summarizes the local quality and adds information about possible late detection. Experiments with real and artificial data have confirmed that we have created flexible tools that can reduce time needed to evaluate detection software efficiently, and provide more detailed information about the quality of detection than the Jaccard index. Their use can significantly speed up data analysis and capture the weaknesses and limitations of the detection system under consideration. Our detection quality assessment method can be of interest to all engineers involved in machine recognition of video data.

Dual-View Single-Shot Multibox Detector at Urban Intersections: Settings and Performance Evaluation

“…Indeed, it may often occur that multiple boxes are predicted for the same ground truth object. The NMS algorithm filters out the predicted boxes based on the class confidence and the intersection over union (IoU) method [ 31 ] between them. In particular, for a given SSD output grid and class, for each real object, the predicted box (if any) with the highest class confidence is picked.…”