Traffic Lights Detection and Recognition Algorithm Based on Multi-feature Fusion

Abstract

Detection and recognition of traffic lights is important for intelligent assisted driving. Traditional color space based traffic lights detection algorithms could be easily affected by other objects (such as buildings, car taillights) in the surrounding environment, and the detection accuracy and real-time performance are not ideal enough. Generally, the deep learning based methods have better advantages of real-time and accuracy performance for the normal scene with obvious traffic lights targets. However, the small traffic lights targets detection rate and accuracy in night-time of these methods are still can't be satisfactory. To solve this problem, this paper proposed a novel traffic lights detection and recognition algorithm based on multi-feature fusion, which can be implemented in two steps (detection and recognition). For the first step, the SLIC (simple linear iterative clustering) super-pixel segmentation algorithm is used for purposes reducing the image data processing complexity and improving the real-time performance. The mean-shift algorithm was used to cluster the HSV (Hue, Saturation, Value) color space components respectively for enhancing the target data and reducing the interference from other targets. For the second step, the feature information extracted by CNN (Convolutional Neural Network) and HOG(Histogram of Oriented Gradient) feature are fused. The SVM (Support Vector Machine) classifier is trained on a data set of traffic lights established by our own. To verify the proposed algorithm in this paper, amount of experiments were carried out in real traffic scenes. Experimental results show that this algorithm almost has the same real-time performance with YOLO_V3 neural network and a better accuracy.