Abstract:
Abstract: In order to monitor illegal land use in real time, video surveillance technology was used to monitor the vulnerable areas of illegal land use. Excavator was one of the most important construction machinery in the engineering construction, an automatic real-time detection of excavator could provide important information for non-contact field monitoring of illegal land. But it was difficult to accurately detect the excavator due to the complex background, uneven illumination and partial occlusion in natural scene, This paper proposed a real-time excavator detection algorithm in natural scene based on the SSD-like (Single Shot Detector). Specifically, the lightweight network DDB (Depthwise Dense Block) was used as the basic network to extract shallow feature and fuse with high-level features in the excavator objection model to enhance the feature representation capability. Meanwhile, the BDM (Bottleneck Down-sampling Module) which was designed based on the lightweight network MobileNetV2 was used as the multi-scale feature extraction network to reduce the parameter quantity and computation. The data sets included 18 537 images of excavators with different shooting angles and natural scenes, 15 009 images were used as training set and 3 528 images were chosen as test set randomly. To enhance the diversity of training data, data set expansion methods such as rotation and image were adopted. Based on the Caffe deep learning framework, the proposed model in this paper was trained with end-to-end approximate joint methods and the model weight was fine-tuned by using SGD (Stochastic Gradient Descent) algorithm. The DDB module of the network was initialized with the weights pre-trained on the PASCAL VOC dataset, and the training time and resources were further reduced by transferring learning. Then the model pre-trained on the large data sets was transplanted to excavator object detection by transfer learning. The proposed method was transplanted and performed on the mainstream Jetson TX1 embedded hardware platform, and experiments on the actual data set of detecting excavator object at different angles of view and natural scenes. Experiment results showed that the parameter quantity and computational complexity of proposed model with BDM was reduced by 68.4% compared to SSD, the mAP (Mean Average Precision) of proposed method reached 90.6% on the testing set, which was 0.4% and 0.4% higher than that of SSD based on VGG16 basic net and MobileNetV2SSD based on MobileNetV2 basic net, respectively. The model size of propose method was 4.2 MB, which was about 1/25 and 1/4 of SSD and mobilenetv2ssd, respectively, and the time-consuming of each frame was 145.2 ms, which was 122.7% and 28.2% faster than SSD and MobileNetV2SSD, respectively. The proposed method not only had better generalization and robustness, but also can be better deployed on the embedded hardware platform which demonstrated the feasibility of real-time monitoring of the excavator at site of illegal land use.