Pedestrian detection is a key problem in many intelligent transport systems. In driving environment, apart from the detection accuracy, the inference speed is also a large concern. Although popular two-stage object detectors such as Faster R-CNN have achieved significant improvements in pedestrian detection accuracy, it is still slow for real-time pedestrian detection in driving environment. On the other hand, popular one-stage object detectors such as SSD have not achieved competitive detection accuracy on pedestrian detection benchmarks. This paper proposes a one-stage detector for real-time pedestrian detection in driving environment. The proposed approach is based on popular SSD framework. To improve the detection accuracy, the backbone network in original SSD framework is replaced by the backbone sub-network based on DenseNets structure, which includes stem module, dense blocks, and transition layers. With dense connection in DenseNet architecture, the proposed approach can achieve higher accuracy with fewer parameters compared with ResNet architecture. In the detection sub-network, enhanced feature extraction subnet takes convolution layers generated by the backbone sub-network to generate enhanced feature maps by fusing operation, atrous convolution and deconvolution operation. Enhanced feature maps can enhance the detection performance of multi-scale pedestrian detection. In addition, residual blocks are added before each prediction layer to reduce the computational cost and improve the detection accuracy. Experimental results on Caltech and CityPersons dataset show that the proposed approach achieves better accuracy compared with popular two-stage detectors while being faster.
展开▼
