A Multi-Objective Recognition Algorithm with Feature Dimension Fusion

Multi-target recognition technology plays an important role in the field of intelligent driving. In this paper, we propose a novel recognition algorithm with feature dimension fusion to further improve the detection accuracy of road targets and accelerate the detection speed of the recognition system. The algorithm proposed in this paper consists of three steps. Firstly, the initial target position is estimated by a full convolutional lightweight network. Then the feature heatmap is obtained by the symmetric feature fusion network. Finally, the initial position estimation is combined with the heatmap to recognize multiple targets. We evaluate our proposed algorithm on the KITTI data set as well as many real scene experiments. Compared to other multi-target recognition methods, our algorithm provides much higher accuracy under different scenes. The experimental results demonstrate that the fusion feature of multi-dimensional feature map can comprehensively describe the overall sensory characteristics of the target, and extend the learning and analysis of neural networks to different levels of vision.