Hardware acceleration for real-time vehicle object detection based on MPSoC *

Abstract

Facing the growing demand for autonomous driving, the accuracy of vehicle target detection has become a crucial issue. Given the challenges of mainstream target detection algorithms on image processors, such as large model parameters, poor real-time performance, high power consumption, and high cost, this paper proposes a real-time vehicle target detection scheme based on MPSoC. By replacing CSPdarknet53 in YOLOV4 with MobilenetV3, lighter feature extraction is achieved. PANET structure is used for feature fusion, and depthwise separable convolution is incorporated. Channel Attention (CA) module is adopted to enhance detection accuracy. Ultimately, the model's floating-point parameters are quantized to 16-bit fixed-point numbers and deployed on the MPSoC. Experimental results on the Kitti vehicle dataset demonstrate that the proposed scheme achieves an precision of 81.97, an FPS of 47.23, and power consumption of 10W, increasing the detection speed by 359.6% compared to CPU and reducing power by 91.4% compared to GPU. The proposed scheme can satisfy the requirements of real-time target detection and can be deployed on power-limited devices such as intelligent vehicles.