Lightweight Convolutional YOLOv5 for Real-Time Object Detection in Infrared Stereo Images

Since under low-illumination and high-interference environments, infrared binocular images still have more accurate depth and target information, how to effectively achieve real-time and accurate object detection on the embedded platform. In this research, we propose an improved YOLOv5-based model which is the fusion of a lightweight convolution backbone network and a multi-scale attention fusion-detection head, which is composed of two parts: a feature extraction stage is designed to introduce separable convolution and group normalization to reduce the number of parameters and accelerate the inference speed; and in the detection stage, a deep-learnable dual-channel attention module (channel attention + spatial attention) is used to fuse and weight the multi-scale features of infrared left and right eye images. At last, with the high detection accuracy, the model complexity and inference delay were greatly decreased, which is applicable to the application scenarios with high requirement for system resources and real-time performance. Experimental results demonstrate that the proposed lightweight convolutional YOLOv5 model can strike a balance between excellent accuracy and real-time performance on resource-constrained platforms for infrared binocular image object detection.