The 20th and 21st joint COTA International Conference of Transportation Professionals (CICTP 2020-21)

Aiming at the problem of low target detection rate and high false alarm rate caused by clutter, noise and device jitter in traditional radar image detection, an improved SSD radar image surface target detection network model TF-SSD is proposed, which is based on multi-scale feature fusion. On this basis, deconvolution network and element sum feature fusion method are introduced to make full use of fusion information between shallow feature map and each feature layer. It not only makes up for the weakness of SSD network which is not robust enough to small targets in radar image, but also reduces network parameters. Experimental results on pre-labeled radar image data sets show that TF-SSD radar target detection network improves 6.2% in comparison with the traditional SSD300 network on mAP and 1.62 in comparison with SSD512 network on FPS, which improves the efficiency of target detection under the condition of ensuring the accuracy of radar image target detection.

CICTP