基于YOLOv3的嵌入式实时视频目标检测算法

深度神经网络在目标检测领域具有优异的检测性能,但其结构复杂、计算量大,难以在嵌入式设备上进行高性能的实时目标检测。针对该问题,提出一种基于YOLOv3的目标检测算法。采用半精度推理策略提高YOLO算法的推理速度,并通过视频运动自适应推理策略充分利用前后帧视频之间目标的关联性,降低深度学习算法的运行频率,进一步提高目标检测速度。在ILSVRC数据集上的实验结果表明,该算法可以在NVIDIA TX2嵌入式平台上实现28 frame/s的视频目标检测,且检测精度与原始的YOLOv3算法相当。

Embedded Real-time Video Object Detection Algorithm Based on YOLOv3

Despite the outstanding performance of deep neural network in object detection,it is hard to implement high-performance real-time object detection on embedded devices due to the complex structure and large amounts of required computation.To address the problem,this paper proposes a YOLOv3-based object detection algorithm.The algorithm uses half precision inference strategy to accelerate the inference of YOLO algorithm.Another inference strategy adaptive to video motions is also adopted to use object correlation between adjacent frames to decrease the running frequency of the deep learning algorithm,and further improve the speed of object detection.Experimental results on the ILSVRC dataset show that the proposed algorithm can implement video object detection on NVIDIA TX2 embedded platforms at a speed of 28 frame/s,and its detection accuracy is close to that of the original YOLOv3.