Motion-aware vehicle detection in driving videos

Abstract

This paper focuses on vehicle detection based on motion features in driving videos. Long-term motion information can assist in driving scenarios since driving is a complicated and dynamic process. The proposed method is a deep learning based model which processes Motion Frame Image. This image merges both spatial (frame) and temporal (motion) information. Hence, the model jointly detects vehicles and their motion from a single image. The trained model on Toyota Motor Europe Motorway Dataset reaches 83% mean average precision (mAP). Our experiments demonstrate that the proposed method has a higher mAP than a tracking-based model. The proposed method runs real-time in driving videos which enables the model to be used in time-critical applications such as autonomous driving and advanced driving assistance systems.

Sda: A Novel Skewed-Deep-Architecture for Vehicle Motion Detection in Driving Videos

“…The values that should not be in the picture but we found are accepted as FP, and the values that should be but we could not find are accepted as FN. When the results were interpreted, a 79% average precision value, which is better than the average precision value obtained on black padded motion profiles(Kilicarslan & Temel, 2022) of the YOLOv3 architecture used for object detection, was found.Mean average precision (mAP) is used to evaluate object detection models. mAP compares the ground truth bounding box with the detected box and returns a score.…”

“…The motion profile shows the relative motion relative to another vehicle. The formula (Kilicarslan & Temel, 2022) for the creation of motion profiles is as in (1).…”