Comparing Detection-Based and Motion-Based Preprocessing of Video Frames for Automated Driving Scenario Detection

In the automotive field, accurate detection and classification of driving scenarios is a key element for safety and efficiency for advanced driving systems. High-level pre-processing is an effective method for accelerating the learning process of neural models by extracting the most informative features from frames while eliminating outliers, noise, and irrelevant data. This research investigates and compares two high-level preprocessing approaches, detection-based and motion-based, in the context of driving scenario video-clip classification. Detection-based preprocessing (DbPP) detect high-level features, such as vehicle bounding boxes and road lane segmentation, to highlight key spatial information. Motion-based preprocessing (MbPP), on the other hand, uses Optical Flow to capture temporal information among frames. While both approaches demonstrate satisfactory performance, our experiments indicate that DbPP generally offers better results and reliability, with an accuracy of 88% compared to MbPP’s 83%. Additionally, DbPP demonstrates faster computation times, making it more suitable for real-time applications, while it requires more memory than MbPP. This study also highlights the significance of effective preprocessing in enhancing the accuracy and efficiency of driving scenario detection.