An object detection approach for lane change and overtake detection from motion profiles
Andrea Benericetti, Niccolò Bellaccini, Henrique Piñeiro Monteagudo, Matteo Simoncini, Francesco Sambo
TL;DR
The paper addresses detecting lane-change and overtaking maneuvers from dashcam footage under edge-computing constraints. It represents driving video as a motion profile, a $W \times T$ image derived by vertically averaging a belt of pixels below the horizon from an input video of size $H \times W \times T$, thereby reducing data from $H \cdot W \cdot T$ to $W \cdot T$. It adapts a YOLOv3-based object detector by replacing standard convolutions with CoordConv layers to provide absolute position information, achieving state-of-the-art results on the motion-profile maneuver-detection task with favorable on-device latency. The work demonstrates practical impact for fleet management and driver monitoring by enabling real-time, lightweight maneuver analysis on edge devices.
Abstract
In the application domain of fleet management and driver monitoring, it is very challenging to obtain relevant driving events and activities from dashcam footage while minimizing the amount of information stored and analyzed. In this paper, we address the identification of overtake and lane change maneuvers with a novel object detection approach applied to motion profiles, a compact representation of driving video footage into a single image. To train and test our model we created an internal dataset of motion profile images obtained from a heterogeneous set of dashcam videos, manually labeled with overtake and lane change maneuvers by the ego-vehicle. In addition to a standard object-detection approach, we show how the inclusion of CoordConvolution layers further improves the model performance, in terms of mAP and F1 score, yielding state-of-the art performance when compared to other baselines from the literature. The extremely low computational requirements of the proposed solution make it especially suitable to run in device.