Event-Based Object Detection with YOLOv5 with Attention

Event-based cameras represent a new paradigm in image processing since they operate based on asynchronous events triggered by light brightness change rather than intensity as in the case of the traditional frame-based cameras. Thereby, along with reductions in power consumption, memory footprint, and latency, they provide High Dynamic Range. However, being a relatively new research field, there are no widely adopted methods to process the data from those sensors. In this work, the performance of a well-recognized frame-based object detection algorithm, YOLOv5, operating over event-based data converted to frames was evaluated within an automotive dataset. Events were converted to frames by sampling the spikes at specific grid locations within a fixed time-window and then used to train the five models from the YOLOv5 family, resulting in mAPs ranging from 0.38 to 0.46, where the state-of-the-art up to that moment registered 0.47. Finally, the runtimes ranged from 7.67 to 15.4ms on a 12GB GTX1080 NVIDIA GPU, being perfectly suitable for real-time operation and being 2.31 x faster than an architecture with similar mAP. In addition, an analysis of the modified YOLOv5 with attention mechanisms was conducted.