On-street parking management is difficult, difficult to operate, less revenue, largely because of the poor front-end recognition effect, especially the traditional high video technology, in its own installation pitch angle, tilt angle large and the complex impact of the surrounding complex environment, there is a series of problems such as missed parking, false alarm anomaly, resulting in the loss of state assets at the same time, but also detrimental to the overall development of the intelligent parking industry.
In order to empower the industry and promote development, Vzneith entered the field of on-street parking in 2019. 2 years, based on 10 years of parking technology accumulation and massive data serving 55% of parking spaces in China, Vzneith has continuously broken through a number of application problems such as three-dimensional chassis regression and ultra-long tracking, and gradually increased the comprehensive recognition rate of the on-street parking high level video technology industry to 95%, empowering the national operators and solving the difficult parking The problem of people’s livelihood.
Three-dimensional chassis regression – vehicle three-dimensional spatial positioning
Based on the IoU value of CNN two-dimensional detection frame and parking frame to carry out vehicle and parking space matching is a major challenge in traditional on-street parking high level video technology.
Because of the actual application, the camera is restricted by the site conditions, need to flexibly adapt to the same side of the parking space, oblique side, opposite side of the imaging perspective and the vehicle oblique parking, vertical parking, side parking and other diverse parking methods, resulting in the installation of tilt angle and pitch angle is too large, resulting in two-dimensional vehicle detection frame and parking frame can not match or misalignment matching.
Theoretically, drawing “algorithmic parking frame” and separately configuring parameters for the scene can improve the matching accuracy, however, this solution increases the difficulty and complexity of construction. For this reason, Vzenith incorporates 3D chassis regression technology.
The technology automatically fits the chassis information of each vehicle by using CNN network to get the three-dimensional occupancy of the vehicle in three-dimensional space, which significantly improves the accuracy of matching the association between vehicle detection frame and parking frame.
It also simplifies the construction process, whether it is the same side, different side, opposite side of the deployment angle, or diagonal parking, draped parking, side parking parking, all only need to draw the actual parking space line, without other parameters configuration, which extends the application scope of high level video technology in on-street parking.
ReID ultra-long tracking-solve 30-minute masking anomalies
State-of-the-art tracking strategy can usually handle object disappearance within 3s only.
In the actual high level camera application, obscuring situations such as 10 seconds (passing car), 10 minutes (traffic jam), or even 2 hours (big car parking) often occur, resulting in the disappearance of the target vehicle, and the resulting missed orders and false alarms in and out of the field are the main obstacles affecting the accurate charging of on-street parking. ZhenSense proposes the technical solution of ReID fusion face recognition and car search by map.
The solution is first based on ReID technology to complete the re-matching of some vehicle components (such as roof, windows, mirrors) and vehicle ID, and join the face recognition and the picture search technology commonly used in traffic management to achieve ReID enhancement to meet the judgment requirements of similar vehicles blocking each other, and have excellent performance in practical applications including opening and closing the trunk, and blocking large vehicles crossing the road.
The solution also passed the very long occlusion dataset test (a total of 115,762 videos in the dataset, containing 10% of the target disappearance data due to occlusion greater than 30 minutes). In the open source solution (GOT) comparison test, which is the highest ranked solution in the MOT challenge website, the MOTA (used to evaluate tracking accuracy: false alarms, missed targets, ID switching, etc.) of GOT is 12.8 and RECALL (rate of correctly matching target detection) is 25.8, while the ZhenSense solution can maintain a MOTA of more than 75 and RECALL of more than 90, which is validated as a solution for effective solution for occlusion.
H series continues to land – accelerating urban parking development
The leapfrog technology upgrade makes ZhenSense on-street parking products quickly gain market recognition after its introduction. Following the expansion of in-street parking low level video pile H1L from Qingdao and Rizhao to Fengqing and other places, high level intelligent camera H1M also rapidly covers more than 20 cities including Chengdu, Mianyang, Chaozhou, Suzhou, etc. in less than 1 year, the complete product matrix of high, medium and low level can adapt to all kinds of in-street parking scenes, which greatly accelerates the speed of landing in the country, and is a strong guarantee for the in-street parking industry to accelerate the upgrade to intelligence and digitalization. It is a powerful guarantee for the on-street parking industry to accelerate the upgrade to intelligence and digitalization.
— Make Intelligent Everywhere —
