An Improved YOLO for Road and Vehicle Target Detection Model

Abstract

The yolo series is the prevalent algorithm for target identification at now. Nevertheless, due to the high real-time, mixed target parity, and obscured target features of vehicle target recognition, missed detection and incorrect detection are common. It enhances the yolo algorithm in order to enhance the network performance of this method while identifying vehicle targets. To properly portray the improvement impact, the yolov4 method is used as the improvement baseline. First, the structure of the DarkNet backbone network is modified, and a more efficient backbone network, FBR-DarkNet, is presented to enhance the effect of feature extraction. In order to better detect obstructed cars, a thin feature layer for focused detection of tiny objects is added to the Neck module to increase the recognition impact. The attention mechanism module CBAM is included to increase the model’s precision and speed of convergence. The lightweight network replaces the MISH function with the H-SWISH function, and the improved algorithm improves by 4.76 percentage points over the original network on the BDD100K data set, with the mAP metrics improving by 8 points, 8 points, and 7 points, respectively, for the car, truck, and bus categories. Compared to other newer and better algorithms, it nevertheless maintains a pretty decent performance. It satisfies the criteria for real-time detection and significantly improves the detection accuracy.