Two-Strea[M] YOLOV8: Object and Motion Detection in Driving Videos

Object detection has numerous applications in intelligent vehicles, as it is crucial to quickly determine an object's location and movement for autonomous driving. Traditionally, most algorithms handle these tasks in sequential steps, detecting objects based on appearance features in video frames, and then analyzing their behavior through frame tracking. This study presents a novel deep learning-based object and motion detection method that uniquely combines spatial and temporal information into a single framework. The motion pattern of objects is uniform across different object classes and appears as traces in the spatial-temporal domain. These object movements can be interpreted from motion profile images even in complex driving environments. Unlike two-stage methods that rely on detection and tracking, our approach directly learns object motion from a vast dataset of driving videos, demonstrating its efficiency and practicality. It is specifically designed to address the challenges encountered in dynamic driving scenarios, proving its effectiveness and relevance in practical applications. The goal is to quickly identify objects and their motion in the driving context. Our method excels in real-time performance with interpretable motion detection in the spatial-temporal domain. It also demonstrates high mean average precision, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\mathbf {78\%}$</tex-math></inline-formula>, and low mean average error, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\mathbf {3.09^\circ }$</tex-math></inline-formula>, on a publicly available dataset, further validating its effectiveness and reliability.