As the core algorithms in artificial intelligence, visual object detection and tracking have been widely utilized in home monitoring scenarios. Recently, IMOU, the smart home brand in China, wins the first places in KITTI 2D object detection of pedestrian, multi-object tracking of pedestrian and car evaluations. It scores 57.15% high-order tracking accuracy (HOTA) for multi-object tracking of pedestrian, 82.08% (HOTA) for multi-object tracking of car, and 82.77% (Moderate) for 2D object detection of pedestrian.
Generic object detection and tracking are fundamental and challenging tasks in computer vision. For 2D object detection, IMOU proposes an algorithm framework upon structural re-parameterization and multi-modal fusion, thus facilitating a significant performance gain when handling occlusion and small objects. This framework has been fully applied to the intelligent cameras of IMOU.
For multi-object tracking, IMOU optimizes object detection, re-identification, object correlation, and other modules based on the TBD framework to improve the overall performance. Specifically, the detection capability is improved based on structural re-parameterization. Besides, we attain a significant performance gain when handling distortion and occlusion by proposing a strategy of dynamic template updating. Furthermore, the graph network is used to model the timing sequences of tracking target to strengthen multi-feature fusion and improve the stability of multi-object tracking. With the help of the proposed object tracking algorithm, IMOU cameras can constantly follow the targets of interest and record the movement trajectory. When an abnormal behavior is detected, our users will receive a warming message.
Taking Cruiser 2, an AI outdoor camera to be launched by IMOU, as an example. Local AI algorithms are equipped with high-powered AI chips to achieve efficient data analysis and operation. The newly upgraded AI functions can accurately recognize pedestrians and vehicles, and constantly follow the targets of interest and record the movement trajectory . In addition, the monitored area can also be intelligently zoned on the app, facilitating the flexibility of smart monitoring.
The KITTI vison benchmark is currently one of the largest evaluation datasets in computer vision. It was jointly founded by the Karlsruhe Institute of Technology in Germany and the Toyota Research Institute in the United States. KITTI is used for the evaluations of stereo vison, optical flow, scene flow, visual odometry, object detection, target tracking, road detection, semantic and instance segmentation. As a provider of full-scenario smart home solutions, IMOU has been working in the field of AI for years and keeps making breakthroughs. On the one hand, our AI technology constantly improves household products in terms of high-level intelligence and humanity. On the other hand, we are trying to change the situation that household products were isolated from each other. In the future, more and more products of IMOU will be connected, facilitating a comprehensive smart home eco-system and creating an easier, smarter and more secure life for every user.
Photo – https://mma.prnewswire.com/media/1983668/1.jpg
Photo – https://mma.prnewswire.com/media/1983669/2.jpg
Photo – https://mma.prnewswire.com/media/1983670/3.jpg
