Real-time pedestrian and vehicle detection on embedded devices play crucial role in many intelligent transport systems because of the limited hardware in autonomous driving devices. This paper presents a lightweight two-stage detector for real-time pedestrian and vehicle detection. The proposed detector includes a lightweight backbone at first stage and a lightweight detection network at second stage. The proposed lightweight backbone is designed based on the ShuffleNetv2 network, which achieves the best accuracy in very limited computational budgets. The proposed lightweight detection network consists of an improved R-CNN to improve the computational cost and a separable convolution module to increase the receptive field. In addition, a lightweight region proposal network is used to improve both accuracy and inference speed of proposals generation stage. The lightweight region proposal network includes pointwise convolution to reduce the number of channels of input features and dilated convolution to enlarge the receptive field. The KITTI dataset is adopted to evaluate the effectiveness of the proposed detector. Experimental results on recent embedded devices, including Raspberry Pi 4 and NVIDIA Jetson TX2, and GPU-based computer show that the proposed method achieves a much better trade-off between accuracy and efficiency compared with recent methods and meets the requirement for real-time object detection on embedded platforms.
展开▼
