Abstract: Deep learning algorithms for tracking small, dim objects in infrared image are widely used in smart security and unmanned inspection. However, challenges arise in deploying these algorithms on compact devices, such as drones and unmanned boats, due to their large model size and computational complexities. These challenges are further exacerbated by constraints on load, space, and power. To overcome these issues, a tracking system for small and dim infrared objects has been developed, utilizing the ZYNQ platform's heterogeneous computing architecture (ARM+FPGA). This system takes full advantage of ZYNQ's programmable hardware and software to optimize the object tracking algorithm network and accelerate neural network parallel computing. The aim is to achieve real-time inference on embedded platforms. A lightweight tracking algorithm, featuring Siamese networks and asymmetric multi-attention modules, has been proposed. This algorithm is quantized to INT8, effectively reducing its complexity and memory access costs. Additionally, a parallelized pipeline circuit, designed according to the computational characteristics of convolutional neural networks and asymmetric multi-attention modules, facilitates accelerated computation. Experimental validation on various hardware platforms using public datasets has shown promising results. The ZYNQ-based system for tracking small and dim infrared objects operates at a low power consumption of 4.23W. It achieves an average computational speed of 125.85 GOPS and a tracking frame rate of 68.07 FPS. These results indicate a significant performance improvement, over nine times that of the i5-9400 CPU.
 

convolutional neural networks; object tracking; hardware accelerator; pipelines; parallel processing systems; real-time; quantization