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

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