SqueezeDet-Based Nighttime Traffic Light Detection with Filtering Rules

Traffic light detection is an indispensable algorithm module in autonomous driving system. In general traffic scenarios, the current mainstream algorithms are able to detect and recognize traffic lights accurately. However, these algorithms may fail in the nighttime detection task due to the quality decrease of camera image, which is caused by the multiple light sources in this scene. Therefore, this paper proposed a SqueezeDet-based nighttime traffic light detection algorithm with false detection filtering rules. The remarkable contributions of this algorithm are: 1) Modifying the anchor size of the native SqueezeDet to fit the bounding box of the traffic lights, which improves the accuracy of the model. 2) Roughly determining the position of the traffic light in the image according to the prior knowledges based on the traffic lights, and the image is cropped to reduce the calculation time of the model 3) Formulating the filtering rules based on the position characteristics of the traffic lights, which improves the precision of the algorithm. In order to verify the performance of the algorithm, we performed experiments on our collected dataset and compared with the advanced target detection technology. The result demonstrates that our algorithm has a significant improvement in accuracy and speed.