PROJECT TITLE :

Small Low-Contrast Target Detection Data-Driven Spatiotemporal Feature Fusion and Implementation

ABSTRACT:

An essential and difficult task in the airspace is the detection of low-contrast targets that are relatively small. In this article, we propose a data-driven support vector machine (SVM)-based spatiotemporal feature fusion detection method suitable for detecting small low-contrast targets. This method is both simple and effective. We come up with a brand new pixel-level feature, which we call a spatiotemporal profile, in order to illustrate the discontinuity of each pixel in both the spatial and temporal domains. The spatiotemporal profile is a local patch of the spatiotemporal feature maps that are concatenated by the spatial feature maps and the temporal feature maps in channelwise. These maps are generated by the morphological black-hat filter and a ghost-free dark-focusing frame difference method, respectively. The spatiotemporal profile is a local patch of the spatiotemporal feature maps. We use labeled spatiotemporal profiles to train an SVM classifier to learn the spatiotemporal feature fusion mechanism automatically rather than the handcrafted feature fusion mechanisms used in previous works. This allows the SVM classifier to learn the spatiotemporal feature fusion mechanism more quickly. The serial SVM classification process on central processing units (CPUs) is reformed as parallel convolution operations on graphics processing units (GPUs), which exhibits a speedup that is over 1000 times greater than the original in our real experiments. This helps to speed up the detection process for high-resolution videos. The blob analysis method is then utilized to complete the detection process and produce final results. Extensive testing is done, and the results of the testing show that the proposed method performs better than 12 baseline methods when it comes to the detection of small low-contrast targets. Based on the results of the field tests, it was determined that the parallel implementation of the proposed method is capable of realizing real-time detection at 15.3 frames per second for videos with a resolution of 2048 by 1536, and that the maximum detection distance for drones in sunny weather can reach one kilometer.