A Lightweight Multi Object Detection Algorithm for Complex Road Scenes Based on CMF-YOLO

By the CMF-YOLO algorithm, it improves both speed and accuracy in identifying objects within intricate driving environments. The algorithm employs an efficient backbone structure, designated as CSCB, which expands the receptive field and encourages inter-channel feature competition. Second, the feature aggregation path MFFP is intended to enhance the aggregation of both shallow and deep features and significantly mitigate the elimination of feature details during the feature fusion process. In addition, the original CIou loss function is optimized to use Focaler-CIoU, which enables the algorithm to focus on targets with difficult and easy samples. This improves the accuracy when identifying fine or occluded targets. A layer adaptive magnitude-based pruning (LAMP) is introduced, which significantly reduces computational complexity. Subsequently, comparative experiments were carried out using the BDD100K dataset. The mAP50 average detection accuracy is improved by 5.3%. The GFLOPs decreased by 43.7%. The number of parameters has decreased by 87.9%. Additionally, the detection speed reaches 135.4 fps. These findings highlight the efficacy of the algorithm in autopilot target detection tasks.