Spilled load detection based on lightweight YOLOv4 trained with easily accessible synthetic dataset

The automated detection of spilled loads on the highway can help quick finding of offending vehicles and prevent traffic congestion. A real-time spilled loads detection algorithm based on the improved You Only Look Once v4 (YOLOv4) is proposed in this paper. Image synthesis method is used to manually generate dataset to solve the problem of insufficient spilled loads images. Six categories of spilled loads are divided into 35,468 samples as a dataset (80% for training and 20% for testing). K-means clustering and Convolutional Block Attention Module (CBAM) were introduced to YOLOv4 Channel pruning was carried out to reduce the computational complexity and resource consumption of the model. The proposed model, of which average precision=98.3%, recall= 0.962, precision=0.98, performs better than other 5 types classifiers and is applied to estimate spilled loads, where the model size and GFLOPs is 134.535 MB and 70 respectively. Compared with the original model, the reduction is 47.6% and 34%, saving computing resources and speeding up the detection speed on the premise of ensuring the detection accuracy. This paper paves a new way for automated spilled load detection with deep learning method, improving the efficiency and accuracy of spilled load detection.