Target Detection of Forward Vehicle Based on Improved SSD
Xiaoying Guo, Liu Qiaoling, Zhikang Qin, Yan Xu · 2021
Vehicle detection plays a vital role in vehicle assisted driving. Aiming at the low accuracy and missed detection problems of SSD algorithm for vehicle detection, an improved SSD algorithm is proposed. In order to extract more vehicle feature information, we propose a SSD network with ResNet50 as the backbone. Aiming at the problem of low recognition rate of small target vehicles, the Feature Fusion Model was designed by fusing the position information of shallow features with the semantic information of deep features; in order to improve the performance of the model, we inserted SE block in the feature extraction layer to analyze different features. The importance of the channel is recalibrated. The experimental results show that the improved method has an average accuracy of 83.09% on the self-made vehicle data set, which is 3.23% higher than the accuracy of the previous algorithm.