Rapid and accurate object detection on drone based embedded devices with dilated, deformable and pyramid convolution

There are more and more demands on deep learning based object detection on embedded devices, such as drone surveillance. However, there are some obstacles for object detection on embedded devices, such as: compu- tational complexity, variance on objects' rotation and scale, and small-size objects. We design an embedded compatible object detection algorithm, which have solve the above problem. We use Dilated and Depth-wise separable convolution to optimize base network for reducing model's parameter and speeding up process. We use Deformable Convolution to sort out variance on rotation and scale. We adopt Feature pyramid structure for locating small-size objects. The embedded platform is NVidia TX2. We have collected data by drone and made a dataset by self. The experiments on our dataset to verify our algorithm. In terms of accuracy, our method achieves high precision on detecting ground objects, while in terms of speed, it processes RGB images of 512 x 512 size in 9 images per second.