Alibaba iDST once again refreshed the world record. This is pedestrian detection and identification.

iDST data icon for KITTI pedestrian detection project

The KITTI algorithm evaluation platform was jointly established by the Karlsruhe Institute of Technology in Germany and the Toyota Institute of Technology in Chicago. It is the largest data collection of computer vision algorithms in the world's largest autopilot scenario for evaluation purposes (motor vehicles, non-motor vehicles). , pedestrians, etc.) Performance of computer vision technology such as detection, target tracking, and road segmentation in an in-vehicle environment.

The image of the Market1501 dataset was collected at a supermarket door of Tsinghua University, and was shot by five high-definition cameras and one low-pixel camera. The data set currently consists of 1501 individuals, consisting of 19,732 gallery images and 12,936 training images. All label boxes are generated by the DPM detector.

In fact, in May of this year, Ali iDST has increased the accuracy of vehicle detection to 90.46%, ranking first in the list. And the latest data on the same difficulty level, Ali iDST has now refreshed to 90.55%.

According to Lei Fengwang AI Technology Review, the technology behind vehicle detection is based on regional fusion decision-making and context-dependent multi-task deep neural network, which is used for vehicle detection tasks in complex scenes, focusing on multi-angle, multi-pose and vehicle occlusion. . In the network structure design, the deconvolution operation is used to improve the recall rate of small targets, and the multi-layer features are spliced ​​to fuse the low-level local information and high-level semantic information, which improves the accuracy of the frame positioning. In the training process, the confrontation training mode in GAN (Generation Against Network) is also borrowed. The team has published many papers in the world's top journals and conferences of computer vision, TIP, ACMMM, etc., and shared their research results.

Pedestrian identification and detection and vehicle detection are both computer vision research content, but their respective difficulties and challenges are somewhat different. Pedestrian detection requires the machine to determine from the image or video whether there are pedestrians and where the pedestrians are; pedestrian recognition requires the machine to recognize all images that appear under a particular camera by a particular person.

Alibaba's iDST deputy dean, IEEEFellow Hua Xiansheng said that most of the images taken by the camera are not visible to the face and need to be identified by the pedestrian's overall and local features. However, in actual situations, including the occlusion, illumination, shooting angle, shooting distance, character posture and other factors, as well as the difference in camera equipment, will bring difficulties to pedestrian identification and detection.

According to reports, Alibaba iDST team proposed a cascade network based on target size classification in pedestrian detection technology, and fully utilized the context information of the region of interest to improve the ability of network feature extraction to solve the problems in pedestrian detection. Problems such as large size fluctuation, occlusion, deformation and inaccurate positioning; at the same time, cross entropy regular constraints are used in target positioning to optimize the positioning accuracy of the frame.

In the aspect of pedestrian re-identification, the team not only uses the latest deep learning technology to extract the global characteristics of pedestrians, but also proposes super-resolution modules and deep attention networks to obtain local details such as head, trunk, limbs, and carriers. A new method of combining coarse-grained global features and fine-grained local features is proposed, which further improves the consistency of pedestrian representation and the accuracy of pedestrian recognition in cross-camera scenarios.

Pedestrian Detection with pedestrian recognition of these two technologies has a wealth of application scenarios, including abortion forecast scenic mall crowd personalized analysis of pedestrian traffic safety, unmanned, finding lost children, the elderly, etc. applications. At present, the above technologies have all been integrated into the brain of Aliyun ET City and have already landed.

Hua Xiansheng said, "As the moon program of the 1960s brought communications technology, bio-engineering technology as a large outbreak, the city brain has become a platform for the world's leading scientific and technological innovation, unprecedented challenges Forced scientists have created an unprecedented technological" .

According to Lei Feng Network AI Technology Review, Alibaba Cloud ET City Brain has landed in Hangzhou, Suzhou, Zhangzhou, Wuzhen and other places. With the help of machine vision algorithms, the city brain of Hangzhou can accurately detect and detect traffic accidents. The average daily alarm is more than 500 times, and the accuracy rate is 92%.