Research on Real-Time Reasoning Based on JetSon TX2 Heterogeneous Acceleration YOLOv4
Qiang Zhang, Wang Yuanyu, Liang Zhu, Jin Zhang, Lin Yu, Lin Dandan · 2021
Aiming at the problem of high CPU resource occupancy rate and large delay on the embedded side when the YOLOv4 target detection algorithm is in model inference. This research proposes an embedded platform based on JetSon TX2, using CPU+GPU heterogeneous mode, deploying a YOLOv4 network model improved by the K-means clustering algorithm, and real-time detection of infrared long-wave images containing 10 types of military targets. Using the advantages of CPU multicore serial, load the weight file of the YOLOv4 model and the cache of the input matrix; GPU core CUDA core accelerates the convolution and clustering of the YOLOv4 model. The experimental results show that two independent experiments of CPU, CPU+GPU on Jetson TX2 platform, the Mean Average Precision of YOLOv4 model inference is 93%; the delay of detecting a single long-wave infrared image is 21267.9ms and 402.4ms respectively; When the target in the video stream, the frame rate of the CPU is less than 0.1FPS, and the frame rate of the GPU can reach 3.2FPS; the average occupancy rate of multi-core CPU resources is 96.3% and 45.0% respectively. It can be seen that when the mAP of the detection target is 93%, the resource consumption and delay of the CPU+GPU heterogeneous mode are better than the CPU architecture, which can provide a reference for the subsequent real-time detection of military targets in the actual combat environment.