基于视频帧间运动估计的无人机图像车辆检测

基于人工智能(AI)芯片搭建轻量化深度神经网络，可以在无人机(UAV)机载端实现视频中车辆目标的自动检测，具有重要的应用前景。为此，提出了一种针对无人机图像车辆目标的检测方法，并在AI芯片上进行部署与测试。方法具体包括:结合无人机图像中车辆目标的尺寸范围，对MobileNet-SSD网络进行裁剪，构建轻量化单帧图像检测器；为解决小目标特性在轻量网络框架下引发的检测性能下降问题，引入帧间运动矢量估计，根据相邻帧信息辅助预测当前帧丢失目标的位置范围，并利用检测结果进行修正，实现丢失目标的再召回。通过对多个数据集进行融合与自动补充标注，搭建了一个高质量的无人机图像车辆目标数据集；同时将方法在基于RK3399芯片计算的嵌入式开发平台上进行实验验证，结果表明:搭建的网络能够显著减少存储资源占用，具有轻量化的特点；同时相比于单帧检测法，引入视频帧间运动估计方法可以有效提高检测精度，并在AI芯片上实现125.3 ms/帧的检测速度。 

Vehicle detection in UAV image based on video interframe motion estimation

The lightweight neural network embedded on artificial intelligence (AI) chips can realize the onboard automatic detection of vehicle objects in unmanned aerial vehicle (UAV) videos, which is important in practical applications. In this paper, a vehicle object detection algorithm in UAV videos is proposed, and then deployed and tested on AI chips. For the proposed detection algorithm, firstly, the MobileNet-SSD network is clipped based on the range of vehicle objects' size in UAV images to construct a lightweight single-frame object detector. Secondly, the interframe motion estimation was introduced to improve the poor detection performance which is usually caused by small object characteristics and lightweight network. Thirdly, the position range of missing objects in the current frame is predicted according to the information of adjacent frames. Finally, the predicted position is corrected by detection results, and the recall of lost objects is realized. Additionally, a high-quality UAV image vehicle dataset was built by fusion and automatic supplementary annotation of multiple datasets. The proposed algorithm is verified on the embedded development platform based on RK3399 chip. The results show that the network with the proposed algorithm can significantly reduce the occupation of storage resources with the lightweight characteristics. Compared to the traditional single-image detection algorithm, the proposed algorithm can effectively improve the detection accuracy. Moreover, detection speed can be as low as 125.3 ms per frame on the AI chip.