Real-Time Vehicle Object Detection Method Based on Multi-Scale Feature Fusion

Abstract

Existing object detection algorithms are affected by scenes with poor robustness, besides the existing public datasets are not applicable to urban road traffic scenes. In order to solve the problems of low accuracy in detecting panoramic video images, high false detection rate, this article designed a real-time traffic information detection method based on multi-scale feature fusion. For a start, the vehicle equipped with hp-f515 driving recorder collected video under the real road scene in Beijing. The total length of the driving route was 11 km. Extracted the recorded video, divided the video into frames which size was 1920 pixel $\times1080$ pixel, the classification was based on the type of vehicles in the general road landscape, and the format of datasets was Pascal VOC. Subsequently, an improved SSD (Single Shot Multi Box Detector) detector was designed, used single-data deformation data amplification methods to perform color gamut transformation and affine change on the original data to generate new data types; utilized the learning rate-adaptive adjustment algorithm to improve the efficiency of detector training. Eventually, the detector was used to detect traffic information in actual road scenes. The experimental results were compared with other traditional detectors. Extensive detection experiments showed that the detector had a processing speed of 55.6ms/frame with an accuracy rate of 98.53%, which could accurately identify multiple objections, small-distant objections, and overlapping objections in actual road scenes. It could provide the advanced driving assistance system with the perception information of the surrounding environment in time.