An Obstacle recognition algorithm Based on Modified SSD
Jinfan Cai, Yahong Ma, Jiaxin Tao, Wendou Nie · 2020
In order to solve the problems of the traditional target detection algorithm in the process of obstacle recognition, such as slow speed, low precision and easy to be affected by complex background. This paper proposes an obstacle recognition method based on improved SSD (Single Shot MultiBox Detector). Firstly, the network structure is optimized. Soft-NMS(softening non-maximum suppression) is used to obtain anchor boxes, providing better initial values for bounding box's prediction. SPP(spatial pyramid pooling) method is used to keep the image scale unchanged and reduce the overfitting. At the same time, batch processing normalization BatchNorm is used to randomly initialize the training model to obtain a stable and predictable gradient, which can stabilize the detector with random initialization training while maintaining good performance independent of the network architecture. Finally, four kinds of obstacles collected in the natural background were identified. The test results showed that the recognition speed of obstacles was the fastest based on the improved SSD, and the average recognition accuracy mAP could reach 79.95%, which was 5.35% higher than that of ssd-300. The SSD method proposed in this paper can effectively detect and recognize obstacles.