Smoky vehicle detection based on multi-feature fusion and ensemble neural networks

Existing methods of smoky vehicle detection from the traffic flow are inefficiency and need a large number of workers. To solve this issue, we propose an automatic smoky vehicle detection method based on multi-feature fusion and ensemble back-propagation neural networks (E-BPNN). In this method, the Vibe background subtraction algorithm and some rules are adopted to detect vehicle objects. To obtain the key region at the back of the vehicle where may be the most possible to have black smoke, the integral projection is utilized to detect the position of the vehicle rear. The proposed LHI features, which fuse Local Binary Pattern (LBP), Histograms of Oriented Gradients (HOG) and Integral Projection (IP), are extracted from the key region. The E-BPNN are adopted to distinguish smoky vehicles and non-smoke vehicles by making classification of the extracted LHI features. The proposed algorithm framework can automatically detect smoky vehicles through analyzing road surveillance videos which obtained in the daytime with good weather conditions. The experimental results show that the proposed method of the E-BPNN with multi-feature fusion has a better performance than the method of the BPNN with single features. In addition, the proposed method also has low false alarm rates than common smoke and fire detection methods.