Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-12-13 Cooperative journals: 《计算机应用研究》
Abstract： Aiming at the problem that the existing target detection framework based on big data and deep learning is difficult to realize real-time video target detection on low-power mobile and embedded devices, this paper improves the target detection framework SSD (single shot multi-box detector) based on deep learning, and puts forward an improved multi-target detection framework LSTM-SSD (long short term memory, LSTM) , which is dedicated to multi-target detection of traffic scenes video. Combining single image detection frame with recursive neural network lstm network to form an interleaved circular convolution structure, the temporal association of network frame-level information is realized by extracting the feature map between propagation frames by adopting a little neck - lstm layer, which greatly reduces the network calculation cost. Combining the time-aware information with the improved dynamic Kalman filtering algorithm, the tracking and identification of the targets which are influenced by strong interference such as light change and large-area occlusion in the video can be realized. Experimental results show that the improved lstm - SSD can achieve good results when dealing with the difficult detection situations such as multi - targets, cluttered background, light changes, fuzziness and large-area occlusion. compared with other target detection frameworks based on deep learning, the average accuracy rate of all kinds of target identification is increased by 5~16 %, the average accuracy rate is increased by 4~10 %, the multi-target detection rate is increased by 4~19 %, and the detection frame rate reaches 43 frames / s, basically meeting the requirements of real - time. The balance between the accuracy of the algorithm and the running speed is achieved, and a good detection and identification effect is achieved.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-08-13 Cooperative journals: 《计算机应用研究》
Abstract： Aiming at the problems that the existing target detection framework based on big data and depth learning has poor recognition effect on low-resolution small targets in high-resolution complex large-field scenes, and the accuracy and real-time performance of multi-target detection are difficult to balance, improve the single shot multi-box detector based on depth learning, and propose an improved multi-target detection framework DRZ - SSD (dynamic region zoom - in, DRZ) , which is dedicated to multi-target detection in complex large traffic scenes. The detection is carried out in a coarse-to-fine strategy, training a low-resolution coarse detector and a high-resolution fine detector respectively, downsampling the high-resolution image to obtain a low-resolution version, designing a dynamic region zoom - in network based on enhanced learning, dynamically enlarging the low-resolution small target region to a high-resolution and then using the fine detector to carry out detection and identification, and detecting the remaining image region by using the coarse detector, so that the detection and identification accuracy of the small target and the improvement effect of the operation efficiency are obvious; Adopting fuzzy threshold method to adjust the adaptive threshold strategy can not only avoid adapting to the data set but also improve the decision-making ability of the model and significantly reduce the detection missed alarm rate and false alarm rate. Experiments show that the improved drz - SSD can achieve good results when dealing with weak targets, multi - targets, cluttered background, occlusion and other difficult detection situations. Through testing on the specified data set, compared with other target detection frameworks based on in-depth learning, the average accuracy rate of various types of target recognition has increased by 4~15 %, the average accuracy rate has increased by 9~16 %, the multi-target detection rate has increased by 13~34 %, and the detection and recognition rate has reached 38 frames / s, realizing the balance between the accuracy of the algorithm and the running rate.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-08-13 Cooperative journals: 《计算机应用研究》
Abstract： In order to realize automatic, universal and accurate identification and guidance of unstructured roads for unmanned equipment in the field environment, propose a road guidance algorithm for field scenes based on graph reasoning model and intelligent optimization. Firstly, segmentting the image into homogeneous superpixel blocks, and fusing multi-features of the superpixel blocks to construct a training set. Improving the traditional laplace support vector machine algorithm, combining the location information of superpixel blocks, dynamically selecting superpixel seed blocks in road areas, and training multi-class classifier regressors of superpixel blocks and consistency regressors of adjacent superpixels; Combining the regression values of two kinds of regressors, constructing the energy function of Markov random field and then using the standard graph cutting algorithm to iteratively obtain the minimized energy function to realize the initial road reasoning segmentation. Combining the initial segmentation results of roads, the objective function is constructed according to the constraints set by people's intuitive perception of roads, and using the differential immune clonal evolution algorithm to intelligently optimize and extract the guide lines of roads. The data collected in Zhushan, Nanjing and DARPA grand challenge database are tested, and comparing the results qualitatively and quantitatively with those of classical algorithms. The results show that the extraction accuracy of the guide line of unstructured roads by this algorithm in the field environment is over 91.79 %, compared with classical algorithms, the detection accuracy is increased by 48.1 % and 35.5 % respectively, and the processing efficiency of the algorithm is increased by 98.6 % and 97.8 % respectively, which balances the real-time performance and accuracy of detection and has a strong application prospect.
Hits
Hits
Downloads
Comment