Research And Application Of Complex Scene Parking-Slot Detection Algorithm Based On Visual Perception

With the exponential increase in the number of private cars,parking friction accidents have also increased.This is due to the complexity of vehicle congestion and road conditions.Therefore,there is an urgent need for a widely applicable parking space recognition technology in today’s society.However,existing parking space recognition systems on the market are expensive and face two main problems: first,it becomes difficult to reduce algorithm dependence on hardware when high computing power hardware chips are monopolized by giant companies;secondly,existing parking space recognition algorithms are either traditional visual algorithms(with poor robustness and difficulty adapting to complex scenes)or based on pure deep learning algorithms(requiring too much hardware computing power and unable to be implemented in largescale applications).To solve these problems,this thesis developed a low-cost,low-power consumption parking space recognition system with high robustness,accuracy and speed using Allwinner T5.The system uses four fisheye surround-view cameras as image input sensors and integrates three functions: 360-degree panoramic monitoring,parking spot target detection and distance measurement.In addition,the system uses Nano Det lightweight object detection model with fast speed and good robustness to obtain information about parking spaces from surround-view images.It also ported MNN deep learning framework that supports parallel inference acceleration for training a car park identification model with AP＿50 reaching 92%.Furthermore,this thesis proposes a distance estimation method based on calibration principles and surround-view images for use in driving training scenarios which can monitor real-time relative position between vehicles and stop lines to determine whether they cross over.To evaluate the performance of our proposed parking space recognition system we conducted onboard testing after porting it onto Allwinner T5 SoC platform designed for automotive use cases.We implemented panoramic monitoring function based on Open GL running on T5 GPU while utilizing MNN so that Nano Det car park target detection network could run smoothly on SoC platform through network quantization optimization techniques along with Open CL acceleration.Experimental results using self-collected datasets show that our proposed parking space detection method has good generalization ability for different types of parking spaces in various scenarios,demonstrating its high practicality and reliability.On the embedded device T5 platform,our proposed parking space recognition system can run at a speed of 11 frames per second(FPS)with an accuracy rate of over 92% for car park identification.In addition,we conducted real vehicle testing specifically designed for driving training scenarios which showed that our proposed method can effectively monitor real-time crossing over stop lines with high accuracy and reliability exceeding 90%.In summary,this thesis developed a low-cost and low-computing power domestic Allwinner T5 platform-based parking space recognition technology with good robustness,real-time performance and accuracy.This technology can be widely used in various scenarios such as vehicle driving training and automatic parking while meeting various functional requirements.