System Profile
The system mainly uses access to multi-channel real-time video streams or reads local files, decodes video frames, and performs detection, tracking, and feature recognition of effective human vehicles and other targets. The vehicle or person that satisfies the capture conditions will be captured. The structure information of vehicle characteristics and pedestrian attributes are obtained, and the structure information of vehicle and pedestrian is transmitted to the data analysis platform through the interface, which provides an effective basis for the search of maps and data mining.
System function
l Multiplex video access decoding
It can access a variety of video stream formats such as H. 264, MPEG4, SVAC, and multiple resolution video images. It can support docking platforms or front-end devices that conform to GB/T28181-2011, and can also access other non-standard front-end devices through SDK. Through multi-card cascading, a single device can achieve up to 64 video images of human car features full function recognition.
l Structured vehicle information
Can be used for vehicle number plates, vehicle brands, vehicle sub-brands, vehicle year, vehicle color, vehicle type, vehicle interior characteristics(sun shield, carton, annual inspection mark, seat belt, phone call, etc.), vehicle exterior features(luggage rack, skylight , spare tires, etc.) and other information for structural description output.
l Pedestrian attribute structure
Pedestrian coat color, coat type, trousers color, trousers type, gender, age group, whether there is a backpack, whether there is a hat and other attribute characteristics of the structural description, can test the human body 18 points of bone key points, And recognize the human body posture and movement characteristics.
l Snapshot
The snapshots of vehicles and pedestrians in the video can be captured. When capturing, one or more snapshots can be selected as clear as possible. The single object is captured only once, without re-capturing. The pedestrian mainly captures the positive features.
l Data uploading
Vehicle and pedestrian pictures and structured feature information captured in multi-channel video streams can be transmitted to the platform or other systems in the form of JSON strings.
Technical characteristics
1, the latest version of GPU architecture is highly optimized for deep learning detection and recognition algorithms.
2. We must use tens of millions of samples to train and self study the convolution neural network structure.
3. It can automatically adapt to multi-angle, different resolution, different lighting conditions of the scene applications, including ordinary surveillance images.
4. It can realize the full feature recognition of vehicle and pedestrian at the same time. The accuracy is high, the capture rate is over 99%, the mistake rate and the replay rate are less than 1%.
5. High concurrency. A single card can support real-time detection and output of 16-channel or more video streams. Cascaded multi-card can realize 64-channel video stream analysis.
6, flexible product form, can cooperate at SDK, software or system level.
