- 期刊
Research on Target Detection Algorithm Based on Lightweight Mobilinet in Assistant Driving System
摘要
Due to the large amount of network parameters and slow detection speed of SSD target detection network, it is not suitable for embedded platform to detect road emergencies in real time. On the basis of the original network, this paper uses mobile net network to replace SSD backbone network, and adds five layers of convolution neural network after it. All adopt deep separable convolution operation, and finally get a new network structure. Conv7, Conv10, Conv11, Conv12, Conv13 and Conv14 are used as effective feature layers to obtain the prediction results. In this paper, 10050 images are used to train the new network and detect pedestrians, vehicles and bicycles on the road. The results show that the parameters of the new network are 7.28 million, the detection speed is 45 FPS, and the average detection accuracy is 79.86%. Compared with the original network, the parameters of the new network are reduced by 70% and the detection speed is doubled. Comprehensive analysis can be applied to embedded platform for real‐time detection of road emergencies.