A Novel Visual Object Detection and Distance Estimation Method for HDR Scenes based on Event Camera
Tianhao Wu, ChengYe Gong, Delei Kong, Shu Xu, Qingjie Liu
Abstract
The event camera is a new type of visual sensor inspired by the biological retina. It can efficiently capture the brightness changes of the scene (called events) in real-time and output a sparse asynchronous event stream with a microsecond resolution. Moreover, it has low latency, low bandwidth, Advantages such as high speed, and High Dynamic Range (HDR). In this article, we propose a novel event camera-based visual target detection and distance estimation method. First, use the DVS event camera to obtain a stable and clear cumulative event image; then, because of the characteristics of the cumulative event image, improve the YOLOv5s network structure to obtain the YOLOv5s (+detection) network so that the objects in the event stream data have better detection accuracy; Next, according to the relationship between the size of different objects in the pixel plane and the real distance, a small MLP network is trained to estimate the distance of the object; finally, in the low-light environment, the use of DVS event camera and RGB camera complete the effect of target detection and distance estimation. Experiments show that our proposed event-based method can achieve high detection and estimation accuracy under different lighting conditions.