摘 要：在视频监控场景中，由于车辆外观的多样性、车辆间的相似性以及应用场景的复杂性，导致算法难以区分不同的车辆目标；同时，实际应用对车辆重识别算法的实时性也提出了要求。为兼顾精度和速度，进行了以下工作：首先，构建了一个场景复杂、含有近百万张车辆图片的数据集；其次，对车辆重识别算法的数据预处理、网络结构、后处理三个方面进行详细实验；最后，对模型进行蒸馏与量化，在保证模型精度的情况下提高模型速度。最终在自研测试集上的重识别精度达到了 93.35% Rank1 和 76.30% mAP，推理速度达到了 400.6 FPS，满足了实际应用需求。
关键词：视频监控；车辆重识别；模型设计
DOI:10.19850/j.cnki.2096-4706.2021.17.023
基金项目：广东省重点领域研发计划项目 （2019B010153002）
中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2021）17-0095-05
Research on Accurate and Efficient Vehicle Re-identification Algorithm for Complex Scenes
JIN Xiaofeng，HUANG Yanjie, XU Tianshi, HUANG Yuheng, HUANG Yuezhen
(Guangzhou GRGBanking Equipment Co., Ltd., Guangzhou 510663, China)
Abstract: In the video surveillance scene, due to the diversity of vehicle appearance, the similarity between vehicles and the complexity of application scene, it is difficult for the algorithm to distinguish different vehicle targets; at the same time, the real-time performance of vehicle re-recognition algorithm is also required in practical application. In order to give consideration to accuracy and speed, the following work is carried out: firstly, a data set with complex scenes and nearly one million vehicle images is constructed; secondly, detailed experiments on data preprocessing, network structure and post-processing of vehicle re-recognition algorithm are carried out; finally, the model is distilled and quantified to improve the model speed while ensuring the accuracy of the model. Eventually, the rerecognition accuracy on the self-developed test set reaches 93.35% Rank1 and 76.30% mAP, and the reasoning speed reaches 400.6 FPS, which meets the needs of practical application.
Keywords: video surveillance; vehicle re-recognition; model design
参考文献：
[1] LIU X C，LIU W，MA H D，et al. Large-scale vehicle reidentification in urban surveillance videos [C]//2016 IEEE International 表 5 蒸馏实验结果 学生网络 教师网络 Rank1 mAP 测试集推理时间（秒） ResNet 34-IBN ResNet 101-IBN-NL 93.46 76.41 249 ResNet34-IBN（直接训练） 89.92 73.98 253 ResNet101-IBN-NL（直接训练） 94.59 79.32 335 Conference on Multimedia and Expo （ICME）.Seattle：IEEE，2016：1-6.
[2] CSURKA G，DANCE C，FAN L，e t a l . V i s u a l Categorization with Bags of Keypoints [C]//Workshop on statistical learning in computer vision，ECCV.2004：1-2.
[3] LOWE D G. Distinctive Image Features from Scale-Invariant Keypoints [J].International journal of computer vision，2004，60（2）： 91-110.
[4] LECUN Y，BOTTOU L，BENGIO Y，et al. Gradient-based learning applied to document recognition [J].Proceedings of the IEEE， 1998，86（11）：2278-2324.
[5] KRIZHEVSKY A，SUTSKEVER I，HINTON G E. Imagenet Classification with Deep Convolutional Neural Networks [J]. Communications of the ACM，2017，60（6）：84-90.
[6] ZHENG Z D，RUAN T，WEI Y C，et al. VehicleNet： Learning Robust Visual Representation for Vehicle Re-Identification [J]. IEEE Transactions on Multimedia，2020，23：2683-2693.
[7] HE B，LI J，ZHAO Y，et al. Part-Regularized NearDuplicate Vehicle Re-Identification [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Long Beach： IEEE，2019：3997-4005.
[8] LIU X，LIU W，MEI T，et al. Provid：Progressive and Multimodal Vehicle Reidentification for Large-scale Urban Surveillance 表 6 量化实验结果 量化策略 模型大小（MB） Rank1 mAP 推理速度（FPS） 显存占用（MB） FP32 144 84.54 63.31 257.7 1441 FP16 47 84.41 63.28 384.6 1141 INT8 26 84.01 61.85 427.3 1083 现代信息科技9月17期.indd 98 2022/1/12 10:40:33 第 17 期 2021.9 99 [J].IEEE Transactions on Multimedia，2017，20（3）：645-658.
[9] LIU H，TIAN Y，YANG Y，et al. Deep Relative Distance Learning：Tell the Difference Between Similar Vehicles [C]// Proceedings of the IEEE conference on computer vision and pattern recognition.Las Vegas：IEEE，2016：2167-2175.
[10] LOU Y，BAI Y，LIU J，et al. Veri-wild：A Large Dataset and a New Method for Vehicle Re-Identification in the Wild [C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Long Beach：IEEE，2019：3235-3243.
[11] ZHONG Z，ZHENG L，CAO D，et al. Re-Ranking Person Re-Identification With K-Reciprocal Encoding [C]//Proceedings of the IEEE conference on computer vision and pattern recognition.Honolulu： IEEE，2017：1318-1327.
[12] LUO H，CHEN W，XU X，et al. An Empirical Study of Vehicle Re-Identification on the AI City Challenge [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville：IEEE，2021：4095-4102.
[13] ZHENG Z，JIANG M，WANG Z，et al. Going Beyond Real Data：A Robust Visual Representation for Vehicle Re-Identification [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops.Seattle：IEEE，2020：598-599.
[14] BABENKO A，LEMPITSKY V. Aggregating Local Deep Features for Image Retrieval [C]//Proceedings of the IEEE international conference on computer vision. Santiago：IEEE，2015：1269-1277.
[15] KALANTIDIS Y，MELLINA C，OSINDERO S. CrossDimensional Weighting for Aggregated Deep Convolutional Features [C]//European conference on computer vision. Cham：Springer， 2016：685-701.
[16] HE K，ZHANG X，REN S，et al. Deep Residual Learning for Image Recognition [C]//Proceedings of the IEEE conference on computer vision and pattern recognition.Las Vegas：IEEE，2016：770- 778.
[17] PAN X，LUO P，SHI J，et al. Two at once：Enhancing Learning and Generalization Capacities via IBN-Net [C]//Proceedings of the European Conference on Computer Vision （ECCV）.2018：464- 479.
[18] HU J，SHEN L，SUN G. Squeeze-and-Excitation Networks [C]//Proceedings of the IEEE conference on computer vision and pattern recognition.Salt Lake City：IEEE，2018：7132-7141.
[19] WANG X，GIRSHICK R，GUPTA A，et al. Non-Local Neural Networks [C]//Proceedings of the IEEE conference on computer vision and pattern recognition. Salt Lake City，IEEE：2018：7794- 7803.
[20] HINTON G，VINYALS O，DEAN J. Distilling the Knowledge in a Neural Network [J/OL].arXiv：1503.02531 [stat.ML]. （2015-05-09）.https://arxiv.org/abs/1503.02531.
[21] NVIDIA. NVIDIA TensorRT [EB/OL].[2021-07-25].https:// developer.nvidia.com/zh-cn/tensorrt.
[22] ZHONG Z，ZHENG L，KANG G，et al. Random Erasing Data Augmentation [J/OL].arXiv：1708.04896 [cs.CV].（2017-08-16）. https://arxiv.org/abs/1708.04896v2.
[23] DENG J，DONG W，SOCHER R，et al. Imagenet：A Large-Scale Hierarchical Image Database [C]//2009 IEEE Conference on Computer Vision and Pattern Recognition.Miami：IEEE，2009：248- 255.
[24] CUBUK E D，ZOPH B，MANE D，et al. Autoaugment： Learning Augmentation Strategies from Data [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach：IEEE，2019：113-123.
[25] HENDRYCKS D，MU N，CUBUK E D，et al. Augmix：A Simple Data Processing Method to Improve Robustness and Uncertainty [J].arXiv：1912.02781 [stat.ML].（2019-12-05）.https://arxiv.org/ abs/1912.02781.
作者简介：金晓峰（1985—），男，汉族，山东潍坊人，广电运通智能安全研究院副院长，高级工程师，博士研究生，研究方向： 计算机视觉、视频大数据。