YOLO-MSLite: Lightweight Multispectral Object Detection Algorithm with Feature Channel-wise Knowledge Distillation for Autonomous Vehicles

The assisted driving system is a key strategy for promoting the growth of science and technology since it improves the driving experience, ensures stable vehicle operation, and protects the lives of drivers. Object detection algorithms, which are the basic technology of the assisted driving system, significantly influence its stability and sensitivity. By fusing visible and infrared pictures, multispectral object detection (MOD) methods have been suggested to improve detection accuracy. Nonetheless, the current approaches for feature-level fusion detection exhibit low detection efficiency. To solve this issue, we present YOLO-MSLite, a lightweight multispectral object recognition technique based on feature-channel-wise knowledge distillation. The technique improves the Conv and C3 modules of the YOLOv5 backbone by introducing group convolution, which decreases the number of parameters while allowing the one-stream network to interact with features. To increase the information selection capabilities of YOLO-MSLite, two-stream, and one-stream models are employed as the teacher and student models, respectively. Experiment findings on several datasets show that YOLO-MSLite achieves the same degree of accuracy as existing state-of-the-art approaches while being lighter in structure and more efficient in detection. The validation findings of the algorithm installed on an embedded platform further reveal that the model gets good detection results and can reach the level of real-time detection.